User Models for Adaptive Hypermedia and Adaptive Educational Systems

Abstract

One distinctive feature of any adaptive system is the user model that represents essential information about each user. This chapter complements other chapters of this book in reviewing user models and user modeling approaches applied in adaptive Web systems. The presentation is structured along three dimensions: what is being modeled, how it is modeled, and how the models are maintained. After a broad overview of the nature of the information presented in these various user models, the chapter focuses on two groups of approaches to user model representation and maintenance: the overlay approach to user model representation and the uncertainty-based approach to user modeling.

Full text

1
User Models for Adaptive Hypermedia
and Adaptive Educational Systems
Peter Brusilovsky1 and Eva Millán2
1School of Information Sciences
University of Pittsburgh, Pittsburgh PA 15260, USA
2 ETSI Informática
University of Malaga
peterb@pitt.edu, eva@lcc.uma.es
Abstract. One distinctive feature of any adaptive system is the user model that
represents essential information about each user. This chapter complements other
chapters of this book in reviewing user models and user modeling approaches
applied in adaptive Web systems. The presentation is structured along three
dimensions: what is being modeled, how it is modeled, and how the models are
maintained. After a broad overview of the nature of the information presented in
these various user models, the chapter focuses on two groups of approaches to user
model representation and maintenance: the overlay approach to user model
representation and the uncertainty-based approach to user modeling.
1.1 Introduction
Adaptive hypermedia and other adaptive Web systems (AWS) belong to the class of user-
adaptive software systems [174]. One distinctive feature of an adaptive system is a user
model. The user model is a representation of information about an individual user that is
essential for an adaptive system to provide the adaptation effect, i.e., to behave differently
for different users. For example, when the user searches for relevant information, the
system can adaptively select and prioritize the most relevant items (see Chapter 6 of this
book [125]). When the user navigates from one item to another, the system can
manipulate the links (e.g., hide, sort, annotate) to provide adaptive navigation support (see
Chapter 8 of this book [21]). When the user reaches a particular page, the system can
present the content adaptively (see Chapter 13 of this book [28]). To create and maintain
an up-to-date user model, an adaptive system collects data for the user model from various
sources that may include implicitly observing user interaction and explicitly requesting

direct input from the user. This process is known as user modeling. User modeling and
adaptation are two sides of the same coin. The amount and the nature of the information
represented in the user model depend to a large extent on the kind of adaptation effect that
the system has to deliver.
As mentioned in the introduction, Chapters 1 to 5 of this book are focused mostly on
the modeling side of personalization, while the remaining chapters focus mostly on the
adaptation side. Chapters 1 and 2 are specifically devoted to user models and user
modeling. Beyond this, user modeling issues are discussed at different levels of detail in
several other chapters. This chapter attempts to complement the remaining chapters in two
ways. First, it provides an overview (a “big picture”) of the user modeling side referring
readers when necessary to additional information in chapters within this book. Second, it
attempts to complement other chapters by presenting aspects that are either not covered in
other chapters or covered insufficiently.
To envision the big picture, this chapter follows Sleeman [175] who suggested
classifying user models by the nature and form of information contained in the model as
well as the methods of working with it. Following his suggestions, we analyze user
models along three layers: what is being modeled (nature), how this information is
represented (structure) and how different kinds of models are maintained (user modeling
approaches). In this book, the overview of user modeling along the first layer, the nature
of the represented information, is provided in section 1.2 of this chapter. This section
serves as a basis for understanding the user modeling problem as a whole (Fig. 1.1).
Fig. 1.1. Three layers for the analysis of user modeling approaches and their coverage in different
chapters of this book. Horizontal dimension represents user features reflected in the models.
The review of the structure and the representation of information in user models (the
second layer) is split between this chapter and Chapter 2 [72]. Together these chapters
provide a detailed overview of the two most important and most elaborate types of user
models, which were originally developed in the fields of information retrieval [107] and
intelligent tutoring systems [158]. Information retrieval and filtering systems attempt to
find documents that are most relevant to user interests and then to order them by
perceived relevance. The user model that typically powers this kind of systems is known

historically as a user profile and represents the user’s interests in terms of keywords or
concepts. Intelligent tutoring systems (ITS), strive to select educational activities and
deliver individual feedback that is most relevant to the user’s level of knowledge. The
user model in ITS is known as a student model and represents mostly the user’s
knowledge of the subject in relation to expert-level domain knowledge. In this book,
section 2.3 in Chapter 2 reviews three main types of user profiles for representing user
interests, while section 1.3 (in this chapter) reviews the dominant overlay approach for
modeling user knowledge.
Finally, the process of construction and maintenance of user models (the third layer) is
discussed in several chapters. The lion’s share of this presentation is provided in Chapters
1 and 2 - the main user modeling chapters. These chapters follow their foci on knowledge
and interest modeling respectively: section 1.4 of Chapter 1 focuses on Bayesian
Networks, the most important approach to knowledge modeling, while section 2.4 of
Chapter 2 reviews major interest-modeling approaches. This presentation is
complemented by several other chapters, which focus on specific user modeling
approaches. Chapter 3 [135] reviews Data Mining approach to user profile constructions.
Chapters 9 and 10 [152; 172] review content-based and collaborative user profiling as
used for recommendation. Chapter 17 [109] elaborates on context modeling for mobile
applications.
The design of this chapter is the result of a compromise between two goals: to provide
core content that is determined by the chapter’s share in presenting the “big picture” of
user modeling while making it useful and meaningful on its own. To achieve the second
goal, the chapter was focused on user modeling for Adaptive Hypermedia (AH) and
Adaptive Educational Systems (AES). AH and AES are two groups of Web-based
systems that extensively employ the user knowledge models covered by the core of this
chapter. To better justify this title, the knowledge modeling core of sections 1.3 and 1.4 of
this chapter were extended to cover overlay modeling and Bayesian user modeling, which
goes “beyond knowledge”. The nature of this chapter caused a different balance between
the breadth and depth of coverage for the different sections. Section 1.2 attempts to
provide a broad coverage while providing relatively few details about specific user models
or modeling approaches. It is intended as a good overview of the topic that will be useful
for anyone interested in adaptive Web systems. In contrast, sections 1.3 and 1.4 provide
more details at the price of a broader coverage. While focusing mainly on the modeling of
user knowledge, they also introduce approaches to the modeling of other kinds of
information about the user.
1.2 What Is Being Modeled
According to the nature of the information that is being modeled in adaptive Web
systems, we can distinguish models that represent features of the user as an individual
from models that represent the current context of the user’s work. The former are

important to all adaptive Web systems while the latter are mostly the concern of mobile
and ubiquitous adaptive systems, where context is essential. This section focuses on the
five most popular and useful features found when viewing the user as an individual: the
user’s knowledge, interests, goals, background, and individual traits. We also discuss
modeling the context of a user’s work. At the end of this section we discuss stereotype-
based user modeling that is an alternative to the more popular feature-based modeling.
1.2.1 Knowledge
The user's knowledge of the subject being taught or the domain represented in hyperspace
appears to be the most important user feature, for existing AES and AHS. In AES, the
knowledge is frequently the only user feature being modeled. In AHS, it is used by the
majority of systems for both adaptive navigation support and adaptive presentation. The
user's knowledge is a changeable feature. The user’s knowledge can both increase
(learning) or decrease (forgetting) from session to session and even within the same
session. This means that an adaptive system relying on user knowledge has to recognize
the changes in the user's knowledge state and update the user model accordingly.
The simplest form of a user knowledge model is the scalar model, which estimates the
level of user domain knowledge by a single value on some scale – quantitative (for
example, a number ranging from 0 to 5) or qualitative (for example, good, average, poor,
none). Scalar models, especially qualitative, are quite similar to stereotype models. The
difference is that scalar knowledge models focus exclusively on user knowledge and are
typically produced by user self-evaluation or objective testing, not by a stereotype-based
modeling mechanism. Despite their simplicity, scalar models can be used effectively to
support simple adaptation techniques in AHS. A number of AHS’s use scalar knowledge
models to support adaptive presentation. These systems divide their users into two or three
classes according to their knowledge level of the subject (i.e., expert, intermediate, and
novice) and serve different versions of the whole page content [63] or page fragments [8;
14; 16] to users with different levels of knowledge.
A good example of adaptive presentation based on a scalar model is the MetaDoc
system [14], MetaDoc represents user knowledge of UNIX (which was the domain of the
system) as a qualitative scalar value (novice - beginner - intermediate – expert. The scalar
model was used to generate an original adaptation based on stretchtext. Stretchtext is a
special kind of hypertext where clicking on an anchor (hotword) simply “expands” it by
inserting a fragment of content after or instead of the anchor. Another click collapses the
expanded content back to the original hotword (a similar approach is used by the well-
known Windows Explorer). Each page in MetaDoc is a stretchtext, which may contain
many expandable hotwords. The idea of adaptive stretchtext presentation in MetaDoc is to
present a requested page with all stretchtext fragments not relevant to the user being
collapsed and all fragments relevant to the user being expanded. To achieve this result, an
author must first classify expandable text fragments as either an additional explanation or
a low-level detail. The user of MetaDoc with an expert level of knowledge of a concept

will be presented with additional explanations hidden (collapsed) and low-level details
expanded. On the other hand, the user with a beginner’s level of knowledge will receive
expanded additional explanations in all cases. After its presentation, a stretchtext page can
be further adapted by the user who is free to expand and collapse text fragments. The
study of MetaDoc demonstrated that this simple technology based on a scalar model can
increase the speed and quality of user comprehension of the content [14]. More
information about stretchtext and other adaptive presentation techniques can be found in
Chapter 13 of this book [28].
The shortcoming of the scalar model is its low precision. User knowledge of any
reasonably-sized domain can be quite different for different parts of the domain. For
example, in word processing, a user may be an expert in using text annotation, but a
novice in formula editing [66]. A scalar model effectively averages the user knowledge of
the domain. For any advanced adaptation technique that has to take into account some
aspect of user knowledge, the scalar model is not sufficient. For the above reason, AES’s
that focus on advancing user knowledge and many AHS’s use various kinds of structural
models. The structural models assume that the body of domain knowledge can be divided
into certain independent fragments. These models attempt to represent user knowledge of
different fragments independently. By the nature of represented knowledge, structural
models can be independently classified along two different sub-dimensions, according to:
• the type of represented knowledge (declarative vs. procedural), and
• a comparison of the user’s knowledge—represented in the model—to an expert’s level
of knowledge of the subject, referred to as domain model, expert model, or “ideal
student” model.
The most popular form of a structural knowledge model is an overlay model. The purpose
of the overlay model is to represent an individual user's knowledge as a subset of the
domain model, which reflects the expert-level knowledge of the subject. For each
fragment of domain knowledge, an overlay model stores some estimation of the user’s
knowledge level of this fragment. The pure overlay model, developed in the field of ITS
over 30 years ago [194], assigns a Boolean value, yes or no, to each fragment, indicating
whether the user knows or does not know this fragment. In this case, user knowledge is
represented at each instant of time as an exact subset or “overlay” of expert knowledge. In
its modern form, an overlay model represents the degree to which the user knows such a
domain fragment. This can be a qualitative measure (good-average-poor), or a quantitative
measure, such as the probability that the user knows the concept.
Since the overlay model represents the user’s knowledge as a (weighted) subset of
expert knowledge, the nature of the user knowledge reflected in the overlay model
depends on the nature of the expert knowledge represented in a specific system. The
majority of ITS focused on representing two types of domain knowledge: conceptual
knowledge (facts and their relationships) and procedural knowledge (problem-solving
skills). Conceptual knowledge is typically represented in the form of a network of
concepts. Procedural knowledge is most frequently represented as a set of problem-
solving rules. Several knowledge-representation approaches, such as ACT-R [2], or
propositional representation [100] allowed one to combine these two types together. More

recently, Stellan Ohlsson suggested focusing on a different kind of procedural knowledge
– not the knowledge that allows the user to solve the problem, but the knowledge that
allows him/her to evaluate the correctness of the solution [144]. This knowledge is
typically represented as a set of constraints.
In turn, the nature of knowledge represented in a specific adaptive system is
determined by the kind of personalized support it provides. A large class of ITS known as
“tutors” focus on helping users solve educational problems and thus rely on the procedural
knowledge of either problem solving or evaluation nature. Other types of ITS and IES
focus on helping users to select the most relevant piece of educational content and thus
rely on conceptual knowledge about the domain. The use of conceptual knowledge is
shared by almost all non-educational AHS, which also focus on guiding the user to the
most appropriate content. In all these cases, regardless of the nature of the represented
expert knowledge, the overlay model can be successfully applied to model individual user
knowledge, i.e., it can measure how well the user knows a concept, what is the probability
that the user can apply a rule or which of the constraints or propositions are likely
mastered.
Overlay models constitute a dramatic step forward from scalar models. Yet, in the field
of ITS, overlay models have often been criticized for being “too simple.” It has been
argued that the state of user knowledge is never an exact subset of expert knowledge. The
user may have misconceptions and her knowledge generally progresses to expert-level
knowledge not by “filling the gaps,” but through a complex process of generalization and
refinement. To model user misconceptions, an overlay model was expanded into a bug
model, representing both correct knowledge and misconceptions (known as buggy
knowledge or “bugs”). Bug models were predominantly used to model user procedural
problem solving knowledge. The most extensively studied form of bug model is called the
perturbation model. This model assumes that several incorrect perturbations can exist for
each element of domain knowledge. Incorrect user behavior may, from the viewpoint of
this approach, be caused by the systematic application of one of the perturbations in place
of the correct rule. The goal of a system with a bug model is not just to declare that a
specific element of domain knowledge is incomplete or missing, but to identify, if
possible, specific buggy knowledge that can be used to provide a higher quality
adaptation. An even richer model that makes it possible to reflect the development
(genesis) of user knowledge from the simple to the complex and from the specific to the
general is known as a genetic model [74].
While both bug models and genetic models are certainly more powerful than the
traditional overlay model, they are also much harder to develop. Research on these models
has contributed to the development of the fields of cognitive modeling and ITS [194], but
the practical use of these models has been quite limited. Genetic models have never been
used in practical systems. Bug models have been used mostly in problem solving ITS
created for simple domains, although several well-known systems created by Carnegie
Mellon researchers have demonstrated that this approach could work for large-scale
practical systems [1; 106]. In the area of Web-based systems the use of bug models is
limited to a small subset of Web-based AES that are focused on adaptive problem solving

support. Non-educational Web systems do not use bug models since they have no means
(and no need) to diagnose misconceptions.
The typical pattern of bug model application in Web-based AES can be demonstrated
by such systems as WITS [145], ILESA [116], and Web-PVT [192]. These systems,
developed for three different educational domains, provide a personalized analysis of
exercise solutions and some form of guidance through problem selection (WITS), problem
generation (ILESA) or adaptive navigation support (Web-PVT). The systems use a
combination of an overlay model and a bug model. Bug models allow the systems to
recognize misconceptions in the user’s problem-solving knowledge, distinguish it from
random slips, such a typos and calculation errors, and provide a useful personalized
explanation. However, the rest of the adaptive functionality (such as problem generation
or adaptive navigation support) only considers the balance between correct and incorrect
usage of knowledge, as reflected in the overlay model, ignoring the exact nature of the
misconceptions represented in the bug model.
In contrast to the powerful but complicated and rarely used bug models, overlay
knowledge models are extremely popular in both Web AES and AHS. Almost every Web
AES and a majority of modern AHS are based on some form of overlay models. Multiple
projects have demonstrated that an overlay model provides a good balance of simplicity
and power. The ability to independently measure user knowledge within different
elements of the domain provides a level of power that is sufficient to run the majority of
the advanced adaptation techniques. Yet, overlay model are relatively easy to develop,
especially for cases with less than 100 domain-knowledge elements. Due to the
importance of overlay models for AES and AHS, this chapter provides a special section
on the development of overlay models.
1.2.2 Interests
User interests always constituted the most important (and typically the only) part of the
user profile in adaptive information retrieval and filtering systems that dealt with large
volumes of information. It is also the focus of user models/profiles in Web recommender
systems. In contrast, early AES and educational AHS paid no attention to user interests,
instead focusing on learning goals when sequencing educational content. As for non-
educational AHS, their nature and the small size of their hyperspaces created no demand
for interest adaptation. This situation has changed dramatically over the last 10 years.
User interests are now competing with user knowledge to become the most important user
feature to be modeled in AHS. The change was caused by the rapid growth of the volume
of information and the growing popularity of several new kinds of information-oriented
AHS such as encyclopedias [84], hypertextual news systems [3], electronic stores [4],
museum guides [142], and information kiosks [65] where access to information is mostly
interest-driven. Following the pioneer attempts mentioned above, interest modeling was
explored in a number of information-oriented AHS. More recently, the abundance of
available content and the growing popularity of the interest-driven constructionist

approach to education have encouraged more attempts to model user interests in
educational AHS.
Starting from the pioneer systems, AHS focused on a new approach for modeling user
interests. The predominant representation of user interests at that time was the weighed
vector of keywords. This approach was used by nearly every adaptive information
retrieval and filtering system and is still the most popular in these areas. More details
about this approach can be found in Chapter 2 [72]. In contrast to this keyword-level
approach, AHS adopted a concept-level approach to user interest modeling where user
interests are represented as a weighed overlay of a concept-level domain model. The
concept overlay approach to user interest modeling is very similar to the overlay
knowledge modeling approach. Overlay user modeling and specific examples of its
application to user interest modeling are presented in more detail in section 1.3 of this
chapter.
Concept-level models of user interests are generally more powerful than keyword-
level models. Concept-level models allow a more accurate representation of interests.
Given a rich domain model, a concept overlay model can separately model different
aspects of user interests. For example, a news personalization system can model user
interests on distinct topics, based in a specific geographical location, and dealing with
specific named entities [96]. An adaptive museum system can separately model interests
in the designer, style, or origin of a jewelry item [143]. In addition, semantic links in the
domain model allow different kinds of interest propagation to compensate for sparsity, a
standard problem of large overlay models.
Powerful, concept-level models of user interests in AHS have been enabled by the
nature of AHS content that is traditionally manually indexed with domain model concepts.
In closed corpus AHS, such as adaptive information kiosks or museum guides, the content
was indexed at the time of system creation. In systems with expandable corpus, such as
adaptive news systems, new content had to be indexed by a provider at the time of its
insertion into the system [3; 96]. The need for manual indexing originally led to the
establishment of two distinctive groups or systems that are able to adapt to user interests.
Closed corpus systems (mostly AHS) used the concept-level interest model, paying the
price of manual indexing. Open corpus systems (mostly information retrieval and filtering
systems) used keyword-level models, but were able to work with an unrestricted corpus of
document due to their ability to process documents automatically. This dichotomy is no
longer clear-cut. A new generation of adaptive information access systems has attempted
to combine concept-level interest modeling with automatic document processing. Many of
these systems are based on automatic document categorization where each document is
automatically assigned to a concept of an existing domain model (such as the Yahoo!
directory or the ACM topic ontology). These approaches are presented in detail in Chapter
2 [72] of this book. More recent systems explore automatic multi-concept indexing, where
a document can be automatically connected with several domain model concepts [43]. In
addition, a new group of hybrid systems attempts to combine concept-level and keyword-
level interest models in one system [54]. A more elaborate discussion on bridging the gap

between closed corpus AHS and open corpus information retrieval and filtering systems is
provided in Chapter 22 of this book [25].
1.2.3 Goals and Tasks
The user's goal or task represents the immediate purpose for a user's work within an
adaptive system. Depending on the kind of system, it can be the goal of the work (in
application systems), an immediate information need (in information access systems), or a
learning goal (in educational systems). In all of these cases, the goal is an answer to the
question "What does the user actually want to achieve?" The user's goal is the most
changeable user feature: it almost always changes from session to session and can often
change several times within one session of work. Early research on adaptation to the
user’s work goal was done in the area of adaptive interfaces and intelligent help systems
[10; 66; 76]. Adaptation to the user’s learning goal was explored by instructional planning
and sequencing systems [18; 123; 124; 197]. Adaptation to the user’s immediate
information need was explored by adaptive information retrieval systems [17]. Among
AHS systems goal modeling techniques were explored in the sub-areas of hypertext-based
help systems [63; 78] and adaptive information access systems [88; 97; 121] where the
ability to adapt to the current user goal is very important.
Goal modeling in modern AHS and AES mostly follows the approaches suggested in
the pioneer research mentioned above. The user’s current goal is usually modeled with a
goal catalog approach, which is somewhat similar to overlay knowledge modeling. The
core of this approach is a pre-defined catalogue of possible user goals or tasks that the
system can recognize. Frequently this catalogue is simply a small set of independent
goals, however, some systems use a more advanced catalog in the form of a goal or task
hierarchy, which is inherited from earlier research on adaptive interfaces and instructional
planning. In a goal hierarchy, relatively stable higher-level goals are progressively
decomposed into subgoals down to the lowest level formed by short-term goals. Most
typically, the system assumes that the user has exactly one goal (or one goal at each level
of the hierarchy) at any moment of work with the system. The job of the user modeling
component is to recognize this goal and to mark it as the current goal in the model. This
fires the adaptation rules that refer to possible user goals specified in the catalogue. The
adaptation rules can, for example, recommend some pages to the user [23], focus user
attention on a subset of the hyperspace [69; 198], or adapt content of the selected page
[88].
The goal/task recognition process is difficult and not precise in general. It is especially
difficult in AH and other Web-based system where the flow of information from the user
(bandwidth) required by user modeling components is thinner than in traditional desktop
systems. Over the years, adaptive Web systems explored a number of approaches to fight
this problem. To start with, a number of practical systems allow the user to specify the
current goal. Typically the user has to select one of the pre-defined goals [69; 88; 151],
although some systems are able to gradually learn how to adapt to a completely new goal

introduced by the user [97; 121]. A different approach to fight the imprecise goal
recognition is to model the user current goal as a probabilistic overlay of the goal
catalogue where for each goal the system maintains the probability that this goal is the
current goal of the user [63; 126]. Finally, several recent projects explored the use of data
mining technologies to identify the current user task in an expected sequence of tasks and
to provide personalized task-level support [87; 94]. More information about the use of
data mining technologies for Web personalization ns provided is Chapter 3 of this book
[135].
A popular example of goal adaptation is provided by the PUSH system [88]. This
system has a small catalogue of user goals and adapts the presentation of each selected
page to the current goal. Depending on the current goal, some parts of the page can be
collapsed using the stretchtext approach explained in subsection 1.2.1 above. The system
attempts to deduce the user goal by observing the user’s actions and shows the assumed
goal to the user in the spirit of a glass-box adaptation. The user can also change the
current goal by selecting a more appropriate one from the catalog.
A different example of goal modeling is a performance support system ADAPTS [23].
ADAPTS deduces the current goal within a goal hierarchy by following the user’s aircraft
maintenance operations. Once the goal is recognized, the system generates a page with the
list of technical manual fragments that are most relevant to the current goal and user level
of knowledge.
1.2.4 Background
The user's background is a common name for a set of features related to the user's
previous experience outside the core domain of a specific Web system (for example, the
core domain of a city guide [36] is a specific city and its objects of interest; the core
domain for a hospital information system [198] is a specific hospital, its objects and
procedures). A range of backgrounds that have been used in adaptive Web systems
includes the user's profession, job responsibilities, experience of work in related areas, and
even specific view on the domain. For example, medical adaptive hypermedia systems can
distinguish two or three categories of users according to their knowledge of medical
terminology and adapt content presentation to the user category by selecting either
medical terms or everyday language to present the same content [8; 178]. Alternatively
these systems can distinguish users by their profession (student, nurse, doctor) which
implies both the level of knowledge and responsibilities [53; 198]. More examples for this
application area can be found in Chapter 15 of this book [34]. Another example is the
categorizing of users by their language ability (i.e., native or non-native speakers),
followed by choosing the appropriate version of the content for them [102]. Background
information is used most frequently for content adaptation, although there are examples of
the use of it within adaptive search [71; 121] and adaptive navigation support [198].
By its nature, user background is similar to the user’s knowledge of the subject (i.e., it
is also mostly a measure of knowledge beyond the core domain area). However, the

representation and handling of user background in adaptive systems is different. Since
detailed information about the background is not necessary, the common way to model
user background is not an overlay, but a simple stereotype model. In addition, user
background typically does not change during work with the system and is nearly
impossible to deduce by simply watching the user work. As a result, user background is
typically provided explicitly, either by the user herself or by some kind of a superior (a
teacher in a college or an administrator at an institution).
1.2.5 Individual Traits
The user’s individual traits is the aggregate name for user features that together define a
user as an individual. Examples are personality traits (e.g., introvert/extravert), cognitive
styles (holist/serialist), cognitive factors (e.g., working memory capacity) and learning
styles. Similar to user background, individual traits are stable features of a user that either
cannot be changed at all, or can be changed only over a long period of time. Unlike user
background, however, individual traits are traditionally extracted not from a simple
interview, but through specially-designed psychological tests. Many researchers agree on
the importance of modeling individual traits and using them for adaptation. While
different kinds of user traits are extensively discussed in psychological literature, current
work on modeling and using individual traits for personalization focuses mostly on two
groups of traits – cognitive styles and leaning styles. These groups are discussed below.
More recently, researchers on adaptive systems started considering individual traits
beyond cognitive and learning styles. For example, CUMAPH system [188] made a
pioneer attempt to build a user profile from lover level cognitive abilities and apply for
adaptively generating page content for the user. Another recent work considered the use of
personality factors in the context of adaptive museum guides [77] (more information
about adaptive museum guides is presented in Chapter 17 of this book [109]).
Cognitive Styles. By cognitive style, researchers typically mean an individually preferred
and habitual approach to organizing and representing information [165]. Research on
cognitive style has long attracted attention on researchers in Web personalization and
related fields such as human information behavior. Professional literature distinguishes a
number of dimensions in which the users cognitive styles may differ: field-
dependent/independent, impulsive/reflective, conceptual/inferential, thematic/relational,
analytic/global [35; 113]. Most popular among Adaptive Hypermedia researchers are
Witkin’s field-dependent/independent [205] and Pask’s holist/serialist [150]. Since, by its
nature, cognitive style influences humans’ ability to access information, the work on
adaptation to user cognitive style was focused more on the navigational side of AHS and
AES. For example, several adaptive hypermedia systems [61; 133; 190; 191]
distinguished field-dependent and field-independent users and provided different

navigation organization, amount of user control, and navigation support tools for these
groups.
A typical scenario for cognitive style adaptation is provided in AES-CS system [190;
191]. AES-CS attempted to adapt to both user knowledge and user cognitive style. A
field-dependency test was used to classify users in two groups: field-dependent and field-
independent. After that a range of system features were adapted to the identified cognitive
style. Field-independent users received an access to the navigation menu to control their
navigation. F ield-dependent users were only able to proceed through the content
sequentially, however they were provided with additional orientation support tools such
as a concept map and a path indicator. Depending on their learning style, the users also
received different instructions, feedback, and contextual organizers. Evaluated against a
static version of content, the AES-CS (with both kinds of adaptation enabled)
demonstrated a significant increase of user performance [191].
So far, the research on the use of cognitive style of adaptation is a mixed-success story.
On one side, a number of studies confirmed that cognitive style affects both search and
browsing behavior [35; 104]. On the other side, few success stories (like AES-CS) on
using cognitive styles for adaptation were reported. A number of projects made a similar
attempt to distinguish users by their cognitive style and match them with a version of a
hypermedia system developed to support this style, but were not able to report any
significant differences against a non-adaptive condition. It is interesting that attempts to
mismatch user styles and system versions (i .e., to match users to version of the system
developed for the alternative style) typically reported significant negative results [133].
Thus, cognitive style remains an important user feature to take into account, but reliable
approaches of adaptation to user cognitive styles are yet to be found.
Learning Styles. Learning styles are typically defined as the way people prefer to learn.
This group of individual traits is close to cognitive style, but more narrow in scope due to
its focus on human learning. The application of learning styles to adaptation is limited to
Web-based AES. After a few pioneer works that presented the idea of adapting Web-
based AES to learning styles and suggested some ways to implement it [33], this direction
of work quickly emerged into arguably the most popular kind of research on individual
traits adaptation on the Web. A number of pre-Web approaches or inventories to classify
and measure learning styles were applied and a number of studies demonstrated
differences between users with different styles in Web context. Most of the work on
learning style adaptation explored content-level adaptation attempting to match users with
a specific learning style to content that should be the most appropriate for this style. This
adaptation may take different forms, such as selecting the most style-relevant version of
content for presentation, ordering content fragments by their relevance to user style or
hiding style-irrelevant content. More recently, these popular approaches to learning style
adaptation were integrated into several adaptation frameworks and authoring tools such as
ACE [179], CAMELEON [111], AHA! [181], and APeLS [47]. For example, a system
developed according to APeLS framework [47] matches the user model with content

metadata in order to select learning objects that are most relevant to the user’s learning
style given certain alternatives provided in the pool of resources.
Despite all this work, there are no proven recipes for the application of learning styles
in adaptation. It is still unclear which aspects of learning style are worth modeling, and
what can be done differently for users with different styles. Dozens of experimental
systems that consider different style inventories and suggest different ideas for adaptation
were reported, but careful studies are rare and success stories are very few. On the
contrary, a number of experimental studies aimed to evaluate the value of treating users
with different learning styles differently concluded without finding any significant
differences. As a whole, the situation is similar to the situation with cognitive styles: the
area of adaptation to individual traits holds a lot of potential, but offer almost no practical
suggestions. To progress in this area, we either need to learn more about the relationships
between user traits and possible interface settings, or develop trait-agnostic techniques
that treat user traits as a black box and attempt adapt to them using case-based and non-
symbolic technologies [73].
1.2.6 Context of Work
Adaptation to the context of the user’s work is a relatively new research direction within
AHS. It was introduced by several pioneer Web-based systems and later expanded into the
area of mobile adaptive systems. Early context-adaptive systems explored mostly platform
adaptation issues. The growing interest to mobile and ubiquitous systems attracted
researchers attention to other dimensions of the context such as user location, physical
environment, social context, and affective state. Context modeling is conceptually
different from modeling of other user features discussed above. Some information
represented in the context models can hardly be considered information about the user in
pure sense. However, context modeling and user modeling are tightly interconnected.
Many user models include context features; similar techniques are used for context and
user modeling; integrated frameworks are being developed for modeling both user context
and user features [80; 210]. For all that reason, the overview of context modeling is
included in this chapter.
User Platform. Since users of the same server-side Web application may use different
equipment at different times, adaptation to the user’s platform becomes an important
issue. Adaptive hypermedia systems explored a range of techniques that might be used to
adapt to such aspects of user platform (computing environment) as hardware, software,
and network bandwidth. The largest stream of work focused on adaptation to the screen
size by either converting pages designed for viewing on desktop Web browsers to mobile
browsers or generating pages differently for desktop and mobile applications. Another
stream focused on media presentation capabilities that are a combination of hardware,
available software, and bandwidth. Over the last several years, the work on platform

adaptation has grown from a pioneer research domain into practice and is now in the focus
of W3C [202]. Most recent attempt to standardize the description and use of platform
capabilities can be found in [105].
It is important to stress that platform-oriented context models are different from the
knowledge, task, and goal models reviewed above. A context is typically described by a
potentially long set of name-value pairs where names indicate parameters (i. e, screen
width, type of pointing device, or presence of movie player) and values specify parameter
values in the current context. This is, however, a raw model of the context. While
adaptation rules can be written directly addressing the raw parameters, this solution is
nowadays neither practical nor scaleable. As a result, we observe an emergence of context
adaptation approach that is somewhat similar to stereotype modeling. I.e., a set of all
possible combinations of name-value pairs is mapped to a smaller set of stereotypes that
are, in turn, used by adaptation rules. For example, an adaptive system can distinguish two
or more platforms types where each type is formed by a specific range of platform
parameters. In this situation, platform recognition mechanism uses parameters of a
specific platform to determine the current type. After that, simple adaptation rules check
current type and perform different actions for different platform types. For example, if the
user accesses the system from a handheld device, the system switches on conversion [112;
115; 206] or generation [15; 149] of presentation for a small screen. If a user platform
cannot show color pictures or the bandwidth is low, the system converts the pictures to
black-and-white or low resolution [166]. If the platform cannot show movies due to the
absence of a movie player or low bandwidth, it can replace the movie with a picture or
remove a link to the movie [93; 95]. More advanced technologies can generate
considerably different interfaces for different platform types [11; 46; 62] and even use
platform limitation to the benefits of user modeling [12]. For example, a Palm Pilot
version of an adaptive news system [12] that is characterized by a small screen and low
bandwidth requires the user to request pages of news story one by one thus sending each
time an implicit relevance feedback to the system.
User Location. Mobile context-adaptive systems naturally focused on adaptation to user
location. The modeling and the use of location is slightly different from other context
elements. Most frequently the location is used not to fire adaptive presentation rules, but
to determine a small subset of nearby objects of interests. This subset defines what should
be presented or recommended to the user. This kind of adaptation was explored by early
context-adaptive systems in several contexts such as museum guides [142], tourist guides
[36], and marine information system [70]. Accordingly, user location is being modeled in
a way that supports determining the nearby objects. Depending on the kind of location
sensing it is typically a coordinate-based or zone-based representation. More information
about location-adaptive systems can be found in Chapter 17 of this book [109].
A Broader View of the Context. More recently, research on mobile and ubiquitous
computing has considerably expanded the notion of context [173]. While there is no

definite agreement about what should be included into the area of context, most of the
work on context adaptation in mobile and ubiquitous computing focuses on a common
core that includes environment and human dimensions [91; 173]. The environment
dimension includes spatio-temporal aspect and physical conditions (light, temperature,
acceleration, pressure, etc.). The human dimension includes personal context (user pulse,
blood pressure, mood, cognitive load, etc.), social context [154], and user task. This may
look confusing to the user modeling researchers who consider user tasks as a part of the
user model, not context. To avoid confusion, it is important to remember that the research
on context modeling is going on in two different research communities that consider
context from two different points of view (Fig. 1.2). From the user-centered view
employed in the user modeling field, the user task is not a part of the context, while the
device itself is. From the device-centered view, which is dominant in the mobile and
ubiquitous computing, a range of parameters characterizing the current state of the user
are, indeed, a part of the device context. This view is presented in Chapter 17 of this book
[109], which provides an extensive discussion of context modeling and reviews its use in
several kinds of adaptive systems.
Fig. 1.2. The dimensions of context from user- and device-centered views
To establish a more objective border between user and context modeling, it is useful to
observe that user modeling is focused mostly on the longer-term properties of the user that
are distilled from observations, while context models attempt to represent the current
features of the user and the environment, mostly read from context and physiological
sensors [91; 173]. At the moment it is hard to classify the approaches to broader context
modeling and adaptation in adaptive Web systems since these systems rarely attempt to
model context beyond the platform and location. However, it seems that the emerging
stereotype approach for platform adaptation can work within a range of other context
parameters. Namely, a system may map a long list of name-value pairs supplied by
sensors into a small set of meaningful pre-defined stereotype contexts. Once the current
stereotype context is recognized, the system can use adaptation rules that address these
stereotypes. For example, a context-aware application on a mobile device may use several
parameters to recognize whether a user is rushing somewhere or not (two context

stereotypes) and depending on the current context stereotype present information about
the user’s flight in different ways [137].
Affective State. An important context dimension that deserves to be mentioned separately
is the user’s affective state. Influenced by the idea of affective computing [156], the
research on modeling using the affective state has become very popular in both user
modeling and ubiquitous computing areas. A sequence of workshops devoted to “affective
modeling” has been held in conjunction with the biannual User Modeling conference
series since 1999. The proceedings of the last workshop in the series demonstrate the
ranges of issues that are being explored now [30]. The methods used to model user
affective state inherited approaches from both research areas mentioned above. While the
methods applied by researchers with ubiquitous and pervasive computing backgrounds
focus mostly on using various sensor input, researchers with user modeling background
explored a range of approaches based on observing user-system interaction. Most
important in the context of this chapter is the small stream of works on modeling the
user’s affective state by observing the user’s Web log data. Despite the relatively low
bandwidth of this source, a number of researchers have demonstrated that it can be used to
detect user motivation, frustration, engagement, and disengagement [7; 9; 38; 193]. It is
also interesting to observe that Bayesian networks emerged as the most popular
technology to process both sensor input and the user action logs into user affective state
[7; 42; 159; 193]. Section 1.4 of this chapter, which focuses on Bayesian approaches to
user modeling, provides some further information on Bayesian affective modeling.
1.2.7 Feature-Based Modeling vs. Stereotype Modeling
Feature-based user modeling reviewed above is currently the dominant user modeling
approach in adaptive Web systems. Feature-based models attempt to model specific
features of individual users such as knowledge, interests goals, etc. During the user’s work
with the system, these features may change, so the goal of feature-based models is to track
and represent an up-to-date state for modeled features. An alternative to the feature-based
modeling is stereotype modeling.
Stereotype user modeling is one of the oldest approaches to user modeling. It was
developed in the works of Elaine Rich [163; 164] over a quarter of century ago and
elaborated in a number of user modeling projects. Stereotype user models attempt to
cluster all possible users of an adaptive system into several groups, called stereotypes. All
users belonging to the same stereotype are treated in the same way by the adaptation
mechanisms. A user in a classical stereotype-based system is represented simply as her
current stereotype (i.e., a group she currently belongs to). Naturally, each stereotype
groups together users with specific mixture of features. However, stereotype modeling
ignores the features and uses the stereotype as a whole. More exactly, the goal of
stereotype modeling is to provide mapping from a specific combination of user features to

one of the stereotypes. After that, only the user current stereotype is used for adaptation.
Any changes in the user’s features are responded to by simply re-assigning a user, if
necessary, to a different stereotype. Elaine Rich discusses extensively when would be the
right moment for this [164].
Stereotypes were extensively used in early adaptive systems between 1989 and 1994. A
good overview of this generation of stereotype-based systems is provided in [101]. More
recently, stereotype models were overshadowed by feature-based models, yet stereotypes
were used in a number of adaptive Web systems [3; 5; 71; 127; 192]. In addition, as was
mentioned above, the techniques developed for stereotype modeling and adaptation are
used beyond classical stereotype modeling to manage low-granularity feature-based
models.
A promising direction for the future application of stereotype models is their use in
combination with feature-based models. One of the most popular combinations is the use
of stereotypes to initialize an individual feature-based model [3; 5; 192]. This approach
allows to avoid a typical “new user” problem in feature-based modeling where effective
adaptation to new user is not possible since user modeling is started “from scratch.” A
good example of this combination is provided by SeAN system [3] for adaptive news
presentation. Starting the work with a new user SeAN attempts to map the set of
demographic features provided by the user (such as age, education, and type of job) into a
set of predefined stereotypes. These stereotypes, in turn, are used to initialize the feature-
based model of user knowledge and interests. Another promising direction for stereotype-
based modeling is its use in combination with group models. Group models are becoming
increasingly popular in Web personalization. More information about group modeling is
provided in Chapter 20 of this book [92].
1.3 The Overlay Approach to User Modeling
This section provides more detailed information about an overlay approach to user
modeling, the approach that is at the same time most important and most popular for AES
and AHS. We start with overlay knowledge modeling, which was the original application
of the overlay approach. After that we review the use of overlay approach for modeling
user interests and discuss a generalization of the overlay approach for modeling other user
features.
1.3.1 Overlay Modeling of User Knowledge
As we mentioned above, the idea of overlay knowledge modeling is to represent an
individual user’s knowledge as a subset of the domain model that resembles expert
knowledge of the subject. Overlay models of user knowledge were introduced and
developed in the field of ITS where overlay models were used mainly by systems with

task sequencing, curriculum sequencing, and instructional planning functionalities. The
popularity of this approach among early AES and AHS systems can be explained by their
strong connection with ITS systems. In fact, a number of early AHS were developed in an
attempt to extend an ITS system with hypertext functionality [8; 26; 75; 155]. The overlay
knowledge models proved to be a good match for the core function of AHS: providing
personalized access to information. As a result, within just a few years these models were
accepted as de-facto standard by almost all educational and many non-educational
adaptive hypermedia systems. This section provides more details on both components
required for overlay knowledge modeling: the domain model and the overlay knowledge
model.
The Domain Model. The heart of the overlay approach to knowledge modeling is a
structured domain model that decomposes the body of knowledge about the domain into a
set of domain knowledge elements. These elements can be named differently in different
systems—concepts, knowledge items, topics, knowledge elements, learning objectives,
learning outcomes, but in all the cases they denote elementary fragments of domain
knowledge or information. In this paper we will be referring to these fragments as
concepts. Though this name is slightly misleading1 it is currently the most popular way to
name domain knowledge elements. Depending on the domain, the application area, and
the choice of the designer, concepts can represent larger or smaller pieces of domain
knowledge: from a relatively large chunk of knowledge (i.e., a topic) down to elementary
facts [143; 171], rules [167], or constraints [134].
The simplest form of domain model is formed by a set of independent (unrelated)
concepts. This kind of model is called a set model or a vector model since the set of
concepts has no internal structure [20]. Even this simple form of overlay modeling
provides a powerful platform for maintaining a detailed picture of user knowledge that
can support a fine grain adaptation in relatively large practical systems [22; 52]. The main
problem of vector models is the lack of connection between concepts. A vector model can
register user knowledge of a specific concept, but this does not help to model user
knowledge of other concepts. As a result, when the number of concepts is large and the
number of observations is not sufficient, the system is only able to predict user knowledge
for a very small fraction of concepts, related to existing observations.
In a more advanced form of domain model concepts are connected to each other thus
allowing some inter-concept inferencing. By its origin, it is possible to distinguish two
main types of connected models. The first, relatively rare type is inspired by the ideas of
instructional planning and used in a number of AES and educational AHS [32; 110; 117;
176]. The model is formed by a tree of educational objectives where larger objectives
(starting with the whole course) are progressively decomposed into smaller objectives.
1 The word “concepts” can cause someone to think that concepts can only represent fragments of
conceptual knowledge. However, a concept is a general name that can denote a fragment of
knowledge of any kind, including procedural knowledge.

The second type is both more general and more popular. In this type of domain model,
concepts can be connected by different kinds of relationships, thus forming an arbitrarily
complex network. This kind of model (known as a network model) was inspired by
research on semantic networks (Fig. 1.3). Network domain models were used in many
AHS and AES, including several development frameworks [6; 24; 50; 83; 148; 160; 169;
183; 189; 199].
Existing systems use network models of various complexity, with one or more types of
links. In educational AHS and AES, the most popular links are prerequisite links between
concepts, representing the fact that one of the related concepts must be learned before
another. Prerequisite links can support several adaptation and user modeling techniques.
In many AHS, prerequisite linkage is the only relationship given between concepts [81;
86; 141; 148; 157; 204]. More advanced educational systems as well as adaptive
information systems favor classical semantic links such as "is-a" and "part-of" [23; 51;
183; 184; 189; 199].
Concept 1
Concept 2
Concept 3
Concept 4
Concept 5
Concept N
10
3
0
2
7
4
Fig. 1.3. A network domain model with a simple numeric overlay user model
All kinds of links between concepts are used to improve the precision of user
modeling. When the user demonstrates a lack of knowledge, links can help to locate the
most likely concepts that will remedy the situation. For example, if the user failed to
answer a question that is based on the knowledge of several concepts, the concept that has
fewer connections with the well-known concepts is the most likely source of problem.
When the user demonstrates the presence of knowledge, links between concepts allow the
propagation of the presence of knowledge beyond direct observation. For example,
observing evidence of knowledge about a concept, the system can deduce the presence of
knowledge for its prerequisite concepts. This approach is known as knowledge
propagation. Different kinds of links may cause different types of knowledge propagation.
An interesting approach for knowledge propagation is used in the AHA! system [48],

which allows AHS authors to define new concept relationships types and their
corresponding propagation semantics.
Recently, research on network domain models have been taken to the next level by
several research teams that have attempted to use more formalized and more elaborated
forms of network domain models inspired by the Semantic Web. The core idea here is to
use formal domain ontologies [56; 58; 103; 120; 185] or domain topic maps [55; 146] in
place of informal network domain models as a basis for overlay knowledge modeling.
More information about the application of Semantic Web approaches in adaptive Web
systems can be found in Chapter 23 of this book [57].
The Overlay Knowledge Model. The most important function of the domain model is to
provide a framework for the representation of the user's domain knowledge using the
overlay knowledge model. The key principle of the overlay model is that for each domain
model concept, individual user knowledge model stores some data that is an estimation of
the user knowledge level of this concept. The overlay model is powerful and flexible
because it can measure independently the user's knowledge of different concepts. In the
simplest (and oldest) form it is a binary value (known – not known) that enables the model
to represent user's knowledge as an overlay of domain knowledge. This pure form of
overlay model was used in several early AHS.
An extension of the pure overlay model is a weighted overlay model that can
distinguish several levels of user's knowledge about each concept. This approach is used
by the majority of modern AHS and AES. There are three popular forms of weighted
overlay models: qualitative, simple numeric, and uncertainty-based that correspond to
three approaches to user modeling, weight propagation, and adaptation. Qualitative
models represent user knowledge of a concept as a qualitative value [22; 148] (for
example, good-average-poor). These models are used mostly by systems with rule-based
user modeling and adaptation components since qualitative values are very easy to update
and use with rules. Simple numeric models use a quantitative value (for example, from 0
to 100) to represent the level of user knowledge [24; 51]. These models are explored by
systems with simple algebraic approaches to knowledge modeling and propagation. The
uncertainty-based models use different forms of uncertainty management such as
Bayesian networks of fuzzy logics to model user knowledge. The user knowledge in these
models are represented in the form dictated by the selected approach – most frequently, a
probability that the user knows the concept [82; 177] or a probability distribution [3]. Due
to its importance, the uncertainty-based approach is presented in detail in section 1.4 of
this chapter.
One extension of the weighted overlay knowledge model is a layered overlay model.
A layered model stores several values to represent user knowledge of each concept.
Layered models were suggested to avoid mixing together the estimations of user
knowledge obtained from different sources. For example, the system may choose to store
separately the levels of user knowledge obtained by direct observation and by weight
propagation [189; 203] or levels of user knowledge corresponding to different levels of

concept mastery [27] such as Bloom’s Taxonomy of Educational Objectives [13]. In these
systems, different layers are maintained separately and mixed only in the process of
adaptation decision making.
From the technical side, overlay knowledge models of individual users are typically
stored as a set of name-value pairs where the names indicate domain model concepts and
values represent the level of knowledge according to the representation formalism used in
a system. Advanced overlay models such as layered models, fuzzy models, and models
with probability distributions are stored as name-aspect-value triples. The details and the
structure of the domain model is stored once in the system and is typically not replicated
for individual users. As a result, the amount of information stored for each individual user
is very small, even in systems with very large domain models.
1.3.2 Overlay Modeling of User Interests
Overlay approach to user interest modeling is very similar to overlay knowledge
modeling. User interests are represented as an overlay of a concept-level model of the
domain that the system covers. In general, domain models used for interest modeling and
those used for knowledge modeling are slightly different in their structure and the size of
concepts. However, this is caused mostly by the difference in system types: most of the
domain models for knowledge modeling are developed for AES while most of the models
for interest modeling are developed for adaptive information systems. This is rather a
tradition than a necessity – a number of AES model user interests [85; 146] and many
information systems model user knowledge [3; 143]. Conceptually domain models used
for interest and knowledge modeling are compatible: most of the models build for
knowledge modeling can be used for interest modeling and vice versa. A number of
sophisticated AHS take an advantage of this compatibility suggesting modeling both user
interests and user knowledge as two separate overlays over the same network of concepts
[3; 143; 146; 171].
The variety of known domain models used for interest tracking can be classified into
three groups that bear analogy to similar models used for knowledge tracking. First group
is formed by systems that track user interests using a vector model: a set of unrelated
concepts. In these systems, each information object (a document or a fact) is associated
with one or more concepts. A demonstrated interest in an object is modeled by increasing
interest in corresponding concept(s). For example interest in a craft object made in a
specific style may increase the level of interest in this style [143]. Despite its simplicity,
the vector model allows a reliable prediction of user interests for items that are related to
concepts with previously registered interests. However, vector models do not allow the
prediction of user interests for previously unexplored concepts, which becomes a serious
problem as the set of concepts grows. Nowadays vector models are mostly replaced by
two kinds of connected models – a taxonomy model and an ontology model. A taxonomy
model is formed by a classification hierarchy of concepts that are frequently called topics,
classes, or categories. Parents and children in this hierarchy are connected by topic-

subtopic relationships. This kind of model is preferred by systems with expandable corpus
such as Web directories or adaptive news systems [3; 128; 187] as well as similar open
corpus systems reviewed in Chapter 2 [72] of this book. A new resource could be easily
integrated into a hierarchy of concepts by attributing it to one of the leaf concepts (this
classification can be done either manually by content provider or automatically). To
increase the power of this model, it is possible to use more than one classification
hierarchy to model different aspects of user interests. However, with some exclusion [96],
this faceted classification is used only in closed corpus systems since it requires
classifying each item along each of the taxonomies.
In an ontology model of interests the concepts form a rich network connected by
different kinds of links. Most typical links are the same as in a network modeling of
knowledge: is-a, part-of, and similarity. This kind of model is preferred by closed corpus
adaptive information systems such as museum or tourist guides [171] or store catalogues
[4]. The ontology model provides a richer representation of the world behind the
information system and allows better interest tracking and propagation. Note that many
adaptive systems that claim to use ontology for interest tracking are really using
taxonomies that could be considered as a simple case of ontology. In contrast, some
systems attempt to use very sophisticated ontologies such as WordNet to track user
interests [118].
Both groups of domain models that represent links between concepts allow inter-
concept interest propagation that is very important to fight sparsity and to increase the
precision of interest tracking. In simple hierarchical domain models interest is typically
propagated from child to parent concepts. For example, observed interest in machine
learning causes increase of user interest in the parent concept artificial intelligence [128].
Richer domain models based on concept networks and “true” ontologies allow broader
interest propagation. Typically these approaches ignore the nature of the links: if the
system registers user interest in a specific concept, it assumes that related concepts in the
domain model are also of some interest to the user [170; 187]. Some recent advanced
approaches attempt to increase the precision of interest propagation using rules that take
into account the structure and the type of connection in the domain model. For example, if
a user is interested in a driver of a specific racing team, the user is probably interested in
the team itself [43].
1.3.3 Generalized Overlay Models
User features beyond knowledge and interests are not formally modeled as overlays of
domain models; however the approaches to modeling such features as goals, backgrounds
and even whole stereotypes bear some reasonable similarity to the overlay approach. The
framework for modeling these features is based on a space of possible characteristics (a
set of possible goals, or a set of possible stereotypes) that is technically similar to a space
of all concepts in structured domain models. To reveal this similarity, [19] suggested
considering generalized domain models and generalized overlay models. A generalized

domain model is a set of aspects where aspects can represent any characteristics of the
user such as domain concepts, domain tasks and goals, and possible stereotypes. A
generalized overlay user model is a set of pairs "aspect-value" in which the value in each
pair can be "true" or "false" (indicating if the user has this characteristic) or some
qualitative or quantitative value (measuring an extent to which user has this
characteristic).
This generalization turned out to be very useful to find deep similarities between
different AHS. Moreover, the ability to draw a similarity between approaches to model a
specific user feature (such as background) and the very well explored overlay modeling
approach could be instrumental for developing of more powerful and elaborated models.
Below we provide some examples of using this similarity by re-considering approaches to
goal and stereotype modeling.
Let’s start with goal modeling. As reviewed in section 1.2.3, a system that is able to
adapt to user goals typically uses a catalogue of user goals or tasks that the system can
recognize. A traditional approach to goal modeling attempts to recognize the current goal
in the catalogued set and mark it as current in the user model. Now let’s consider the set
of all goals to be a generalized domain model for goal modeling. As in traditional domain
models, the set of goals can be unrelated or can form some connected structure – such as
goal-subgoal hierarchy. In this context, the generalized overlay model of individual user
can represent a probability that each specific goal in the generalized model is the current
user goal. Now the traditional way of goal modeling by marking the current goal becomes
simply a very primitive case of this generalized goal model (current goal has overlay
value 1 and the others zero). Using this framework in full power allows building more
elaborated goal models that can reflect uncertainty in modeling goals. In addition,
recognizing the similarity with overlay knowledge modeling allows to re-use powerful
uncertainty-based modeling and propagation techniques. Indeed, a few adaptive systems
use some elements of this generalized vision to provide a more elaborated goal modeling.
For example, some systems maintain a hierarchy of possible goals or tasks [198] and a
few pioneer systems use probabilistic overlay modeling where each possible goal in the
catalogue is represented by the system as a probability that it is the current user goal [63;
78; 126].
A similar generalization can be used for representing user stereotypes, backgrounds, or
individual traits. For example, a stereotype user model is based on a collection of possible
user stereotypes. A typical stereotype-based system attempts to recognize current
stereotype, mark it as current in the user model and use for adaptation. An application of a
generalized overlay approach allows us to treat a set of all possible stereotypes as a
generalized domain model. A generalized overlay model now allows handling uncertainty
of stereotype recognition as well as to represent more sophisticated cases of stereotype
modeling such as the ability of the user to belong to several stereotypes at the same time.
With this generalization, stereotype user models can be represented as a set of pairs
"stereotype-value,” where the value of each pair can either be "true" or "false" (which
means that the user belongs or does not belong to the stereotype) or some probabilistic
value (which represents the probability that the user belongs to the stereotype). A good

case for maintaining a probabilistic overlay for modeling user stereotypes is provided by
SeAN system [3] that uses a popular combination of stereotype and overlay modeling.
SeAN attempts to map the set of demographic features provided by the user into a set of
predefined stereotypes that are used to seed the overlay model of user knowledge and
interests. Unlike typical stereotype modeling systems, SeAN is not trying to come up with
the single, best stereotype for each user, but maintains a probabilistic overlay. This helps
the system to achieve a more precise seeding of the overlay model.
The ability to view models of different user characteristics as generalized overlay
models allows developers to better understand the similarity between different systems
and to reuse more broadly the representations and user modeling approaches developed in
the field. For example, as demonstrated in the next section of this chapter, elaborated
approaches for knowledge modeling based on Bayesian networks could applied for
modeling other user features such as individual traits.
On the practical side, the vision of generalized overlay models is currently supported
by some universal authoring frameworks for developing AHS. The best example of such a
universal framework is AHA! [49; 50]. AHA! allows the author of an AHS to introduce
generalized concepts and relationships between the concepts. The concepts in AHA! can
represent different user aspects such as knowledge, goals, interest, etc. For each concept,
the system maintains an individual numeric overlay value. The semantics of the concepts,
the connections, and the meaning of specific overlay values are encoded by the author into
a set of user modeling rules and weight propagation rules. The universal nature of AHA!
made it the most popular authoring tool in the field of adaptive hypermedia. So far, AHA!
has already been used for modeling such user features as knowledge, interests, and even
learning styles [181].
1.4 Uncertainty-based User Modeling for Adaptive Hypermedia and
Adaptive Educational Systems
There is no doubt that in the case of user modeling, there is often the need to deal with
information that is uncertain (we are not sure that the available information is absolutely
true) and/or imprecise (the values handled are not completely defined). An example of a
statement that we need to deal with would be: “the user failed this question, so most
probably he/she doesn’t know concept C,” which is uncertain information. On the other
hand, if we say, “the user has been reading about concept C for quite a long time” we are
making an imprecise observation. Obviously, user modeling is a domain in which there
are many different sources of uncertainty and/or imprecision, therefore numerically-
approximate reasoning techniques are suitable for this purpose. The two more commonly
used in this context are Bayesian Networks and Fuzzy Logic.
Bayesian Networks (BNs) were developed in the eighties by Judea Pearl [153] and
since then there has been an increasing interest and enormous progress in the development
of new techniques and algorithms, extensions to the model, and applications. BNs are a

probabilistic model inspired by causality and provide a graphical model in which each
node represents a variable and each link represents a causal influence relationship.
Currently they are considered one of the best techniques available for diagnosis and
classification problems. The Fuzzy Logic (FL) paradigm was originally proposed by
Professor Lofti Zadeh in 1965 [207], and the basic idea is to allow membership functions
(in Fuzzy Sets theory) or truth values (in Fuzzy Logic) to take values between 0 and 1,
with 0 representing absolute falseness and 1 absolute truth. As in BNs, FL has been the
object of intense research, both in the development of models and in its application to real
problems, with special emphasis on control problems. Therefore, it is no surprise that in
the field of adaptive systems researchers have used such models, since the development of
an adaptive system involves diagnosis, classification, and control.
However, in the field of adaptive hypermedia there are few studies that report the use
of approximate reasoning techniques (notable exceptions are [82] for BNs and [99] for
FL). Probably, the reason is that researchers are devoting much more attention to other
relevant aspects such as foundation and core techniques (data mining, meta-data, semantic
web, intelligent web agents, web services, etc.) or practical aspects (architectures, privacy,
security, usability, evaluation, etc). The situation might be comparable to what happened
in other related fields ten years ago (for example, in the field of Intelligent Tutoring
Systems only four of the papers presented at ITS’1996 reported the use of such techniques
(all of them used BNs), while at ITS’2006, eight papers use BNs and one paper uses FL.
Our prediction is that in the future the number of adaptive web-based systems that use
BNs or FL will continue to grow.
Nowadays, most researchers who decide to use approximate reasoning techniques in
their user models choose the Bayesian network paradigm. The reason for this is the
suitability for diagnosis. Also, the fast development of this field over the last twenty years
has resulted in an increasing range of techniques and tools for Bayesian reasoning. In this
sense, tools like GeNIe (SMILE), HUGIN, or JAVABAYES allow for the easy definition
of Bayesian models and seamless integration of learning and updating algorithms. Still,
there are also a number of systems that use other approaches like Fuzzy Logic [37; 98;
108] and, more recently, neuro-fuzzy approaches [182]. The interested reader can see [90]
for an excellent review of former systems and issues in the field of user modeling, and
[99] for some discussion about fuzzy user modeling in the context of adaptive web-based
applications.
At the time of this writing, most solid experience in Bayesian user modeling has been
accumulated in the area of modeling users that interact with educational systems, also
known as student modeling. Student modeling was always the most area for AES and
AHS and it was also the area where new technologies were developed and explored before
being applied for other types of user modeling. Therefore, the core part of this chapter is
focused on Bayesian student models (BSMs), beginning with the modeling of user
knowledge. We will then briefly discuss the use of BNs beyond modeling knowledge. Our
goal is to explain in detail how to define a Bayesian student model (BSM), which student
features can be modeled, and how the parameters of the model can be obtained. The use of
BSM in adaptive web-based systems is still quite unusual. Because of this, we will also

refer to more relevant models that have been deployed in stand-alone tutoring systems and
learning environments. We do this to illustrate how such techniques might be employed in
adaptive web-based educational systems following the success of stand-alone
implementations.
1.4.1 Basics of BNs
A Bayesian Network can be formally defined as:
• A set V of propositional variables (X1, …, Xn) (nodes of the network)
• A set E of probabilistic relationships between the variables (arcs of the network)
Such that:
• The graph G = (V,E) is an acyclic directed graph.
• The conditional independence assumptions are satisfied, i.e., each node Xi is
conditionally independent from the rest of the nodes (except from its descendants),
given the state of its parents.
With respect to nomenclature, the words “node” and “variable” are used interchangeably
in literature. The probabilistic relationship between variables is usually referred to as an
“arc” or “link.”
Though the relationship between the variables does not need to be causal in nature, the
BN paradigm is very suitable for that case. Therefore, building a model usually involves
putting the information available into a cause-effect schema. Regarding the probability
distribution, it can be easily shown that if the conditional independence assumptions are
satisfied, then the joint probability distribution is given as the product of the probability
distribution of each node given its parents. Because of that, in the case of BN it is enough
to provide the conditional probability distribution of each node given its parents (or prior
probability distribution for root nodes) to make any kind of inference needed. In that
sense, a powerful feature of BNs is that they allow for diagnosis (inferences about
possible causes of an event) and prediction (future state/evolution of variables given
evidence).
To illustrate the concepts presented above, let us consider the following example in the
context of student modeling. In order to determine if a student knows a certain domain
element K, we can use the result of a certain event E that provides evidence about it. This
evidence-bearing event can be an answer to a test item, solution of a problem, teacher’s
opinion, the number of Web pages relevant to the element K that have been visited, etc. In
that case, the nodes of the network are: node K (having knowledge about a domain
element K), and node E (result of the evidential event E). In the simplest case, both
variables are binary: K can be known or not_known, while the evidence provided by E can
be positive or negative. To simplify the notation, we will denote the positive states
(known, positive) by 1 and the negative states (not_known, negative) by 0. There is only
one relationship: the state of variable K influences (in this case, causally) the result of the
event E (Fig. 1.4).

K
E
Fig. 1.4. The simplest BN for knowledge diagnosis
Let’s concretize our example. A sample concept could be “Adding natural numbers”
and its possible values could be known and not_known. A possible question could be
“3+4” which has the possible states right and wrong (Figure 1.5).
Add
3 + 4
Fig. 1.5. A simple BN for diagnosing knowledge about adding natural numbers
Regarding the quantitative information (network parameters), we need to provide the
conditional probability distribution of each node given its parents (for the nodes without
parents, the a-priori probability distribution). So in the simplest general case, three
probabilities need to be specified: P(K=1), P(E=1|K=1) and P(E=1|K=0). Let us use the
following values for these parameters:
• P(K=1) = 0.2, means that, in our experience (for instance, if we are a math teacher), the
probability that a random student knows how to add integers is 0.2.
• P(E=1|K=1) = 0.99, means that the probability of giving a correct answer to “3+4” is
0.99 for a student who knows how to add integers. Conversely, although the student
knows how to add, there is still the probability of 0.01 of giving an incorrect answer,
allowing us to model what is called a “slip” in related literature.
• P(E=1|K=0) = 0.02, means that the probability of giving a correct answer to “3+4” is
0.02, if the student does not know how to add fractions. Conversely, even if a student
does not know how to add there is still a certain probability of guessing the correct
answer.
In the related fields of medical diagnosis and statistics such parameters have a clear
meaning. For example, in the medical diagnosis domain, if D represents an disease and T
a test used to diagnose it, the causal relationship is D  E, and the parameters needed are:
the a priori probability of the disease (which in medicine is called the prevalence of the
disease), and the conditional probability distribution P(T/D), more concretely
P(T=1/D=0), which is the rate of false positives (or type II error in statistics) and
P(T=0/D=1), which is the rate of false negatives (or type I error in statistics) of test T.
Both measures are commonly used as indicators of the quality of the test T or its
suitability for the disease D.
Of course, the more complex the structure of the network, the higher the number of
parameters needed. In general, if a node N has n parent nodes P1, …, Pn and each of them
has k states, the number of parameters needed for node N is kn, i.e., the number of
parameters needed grows exponentially with the number of parents.

Once the network has been defined, it can be used to make inferences about the
domain. Using a BN model allows us to reason both in the diagnostic direction (what are
the more probable causes given certain evidences) and in the prediction direction (what is
the probability that a given configuration of variables states will happen, given a set of
evidence?). To illustrate this, we will add some more nodes to our example (Fig. 1.6).
Add
3 + 4
Multiply
3*2 + 7
5 * 4
Fig. 1.6. Adding nodes to the basic BN
Let us imagine that our student gave a correct answer to “3+4.” We can then reason in
the diagnostic direction (i.e., compute the probability that he/she knows concept “Add”)
and also to predict a future result (i.e., compute the probability that he/she will be able to
correctly solve “3*2+7”). This potential for doing both abductive and predictive reasoning
is very useful in the case of student modeling. Not only does it allows us to accurately
diagnose the student’s current state of knowledge, but also to make an informed and
adaptive recommendation for the most appropriate next activity (e.g. exercise, test)
according to the evidence available about this particular student (previous activities with
the system, favorite learning strategy, etc.). For example, for a student that likes to be
challenged it might be better to select a “difficult” (lower probability of right answer)
exercise while for a student that has low self-confidence it might be better to select an
“easy” (higher probability of right answer) exercise.
As pointed out before, BN formalism implements certain independence assumptions
amongst variables. Conditional independence modeled in a particular BN model should
correspond to independence relationships existing in the real world. In our example, the
conditional independence assumptions mean that the knowledge nodes “Add” and
“Multiply” are independent a priori (which is reasonable, because if they are not the
model should include a link between) and that each evidential node is conditionally
independent of the rest of the evidential nodes given the values of its parents (which is
also reasonable, because once the value of the corresponding parent nodes, i.e.,
knowledge nodes, is known, other evidential nodes do not provide information any more).
For more information about what the conditional independence assumptions mean, see
[140] or [153].
The construction of a student model based on BNs consists of two steps: (a)
development of the qualitative model, that involves the definition of the structural model
(nodes and arcs); and (b) development of the quantitative model, that involves the
specification of the parameters needed (conditional and prior probability or distributions).
There are two ways of obtaining the information needed in the above steps. The first way
is to use domain expert’s opinion to get the structure/parameters. The second is to infer

them from data using BNs learning algorithms [79]. A combination of these two ways is
also possible and often employed in practice. As far as learning is concerned, structural
learning is a process of learning dependencies between the given variables (step a), while
parametric learning is a process of learning the parameters describing the strength of the
dependencies in the model (step b).
If a BN is built by an expert, it is necessary to make sure that the model includes all
relevant aspects of the real world, i.e., the model has to account for the relevant variables
and relationships between them, and the specified conditional probabilities have to
represent the strength of these relationships. On the other hand, if a BN is to be learned
from data, the relationships learned on the significance level fixed for the statistical
independence tests performed during structural learning. If the significance level is low
we might miss some weak relationships while if the significance level is high some hard-
to-interpret relationships might be inferred or assumed. A possible solution is to have an
expert to specify the structure (or fine-tune the learned structure) and then learn
parameters from data.
The structure of the rest of section 1.4 is as follows: In subsection 1.4.2, we will
discuss the development of the quantitative model under the assumption that the structure
of a BN is to be built by an expert. To this end, we will discuss which aspects of the real
world being modeled (in our case, a student) can be taken into account and go through
different modeling options2. Subsection 1.4.3 gives some clues for defining qualitative
modeling, while subsection 1.4.4 discusses the combination of expert knowledge and
learning techniques in building BN student models. Section we present some conclusions.
1.4.2 Development of the Qualitative Model
As in any modeling process, the first important step is to specify aspects of the world
being modeled which should be represented in the model, in order to achieve the defined
goals. In the case of student modeling, the range of features to be modeled is wide and can
be grouped into two main categories: knowledge (declarative, procedural, based on skills,
competencies, etc.) and user features (learning styles, cognitive and meta-cognitive skills,
emotional states, etc). The more user features are represented in the model the more
complete it will be. However, the cost of creating and maintaining a more complete model
should be carefully balanced against the usefulness of such a model. In the user modeling
field, the goal is usually to have a personalized model of the user, in order to provide the
basis for adaptation. So we must carefully choose the user features that will be useful to
provide adaptation, because if we are going to consider k different variables and each
variable can take n different values, then we will need to provide adaptive strategies for
the kn different profiles of students generated by the combination of such values.
To begin with, we will explain how to use BNs to model student’s knowledge based on
overlay models. Then we will present how to build knowledge models for other
2 A good tutorial about modeling with BNs for troubleshooting problems is [114].

approaches such us constraint-based models and misconceptions models. Next, we will
discuss how to model features beyond knowledge, such as meta-cognitive skills,
personality traits, affective states, and attitudes and perceptions. Finally, we mention some
examples where the qualitative model has been developed by experts and some other
examples in which learning techniques have been used to infer it from available data.
Along the discussion we will present examples of application, which will illustrate the
main ideas and concepts.
Modeling Knowledge Based on Overlay Models with Bayesian Networks. As
explained before, the first step consists in selecting the variables of interest for our model,
which in this case are the knowledge variables and the sources of evidence that will be
used to obtain relevant information about them. Each one of them will be modeled as a
random variable, and represented as node in the BN. In the case of user modeling,
typically the sources of evidence are the results of interaction with the system (e.g.
answers to questions or exercises, time spent reading certain content, number of clicks,
etc.). Open/scrutable user models allow that either the user or the expert can see and
modify this information.
Once the relevant aspects of the real world being modeled have been selected, it is
time to translate them to a mathematical model. In the case of BNs, this means that we
need to structure this information in a causal relationship schema. To this end, a couple of
important aspects related to building BN models should be clarified: first, that each node
must represent a propositional value; and second, that the proper direction of the links
needs to be established. We will discuss each of these aspects in more detail.
As aforementioned, each node must represent a propositional variable [140], which is
defined to be as a variable that takes an exhaustive and mutually exclusive number of
values. This aspect allows us to determine whether a given element of the real world
should be modeled as a variable or as a state of a variable. Let us illustrate this using some
examples.
Example 1: A medical diagnosis problem
A doctor is trying to diagnose a patient to determine the illness (or illnesses) that he/she is
suffering from. Let us define a finite set of illnesses as {I1, …, In}. In this case there are
two modeling options: (A) to consider a node I with values I1, …, In, and (B) consists in
considering n binary nodes I1, …, In. In the case of medical diagnosis, the choice depends
on whether or not more than one illness can be present at the same time. If the patient can
have one and only one illness, the correct option is A. If the patient can have two or more
illnesses at the same time, the correct option is B. It is important to point out that in option
A all the possible illnesses must be considered within the set of possible values {I1, …,
In}, and if this is not possible, one of the states should be labeled “other” to account for
any illness not taken into account in the model (thus making the set of states exhaustive).
In option B, a node "other" can be included, but it is not necessary.

Example 2: Classification problems
We have to classify a set of objects as belonging to certain categories C1, …, Cn. Again,
the decision of using a single node C with values C1, …, Cn (option A) or n separate
binary nodes C1, …, Cn (option B). The decision depends on whether or not an object can
belong to more than one category. So if we are classifying animals in mammals,
amphibians, reptiles, etc. we should choose option A (with the additional value “others” if
needed). However, if we are classifying persons according to their educational
background into doctors, architects, lawyers, etc. we should use option B.
Example 3. The use of the two different options in user modeling
This example illustrates how the two different options presented above have been used in
two real systems: KBS Hyperbook [82] and the English tutor I-Peter [161]. In both
systems, users are being classified according to their knowledge level K into the
categories novice (N), beginner (B), intermediate (I), and advanced (A). But while KBS
Hyperbook models student’s competence in each knowledge item i by using a node Ki that
takes values {N, B, I, A}, I-Peter models student’s competence in English by using four
separate binary nodes N, B, I, A. In the second system, the underlying assumption is that a
student can belong to more than one category simultaneously. A possible interpretation is
that the I-Peter model might be intended to allow smooth transitions between categories.
For instance, a possible interpretation of P(N=Yes)=0.2 and P(B=Yes)=0.6 is that the user
is gradually leaving the category novice to get to the more advanced state beginner (quite
like in a Fuzzy Logic model). But other combinations of probability values might be
harder to interpret, and therefore estimating the parameters needed might be difficult in I-
Peter.
C2
E
C1
C2
E
C1
Fig. 1.7. A BN and the equivalent BN with the arc C1 E reversed
As already mentioned, a second important decision is the more appropriate direction
for the arcs in our model, i.e., given two propositional variables X and Y, we need to
decide if we will represent the relationship between them as X  Y or X  Y. This is an
especially important aspect, because in spite of the fact that the direction of any arc in a
BN can always be reversed [89], BNs are very suitable for the case of causal relationships
and therefore the model should stay as close to the notion of causality as possible. It is
important to note that if the direction of an arc is reversed new arcs are introduced to the

model. This is necessary for the new network to keep the same dependency structure. Fig.
1.7 shows the arc reversal result on a simple network:
Any of these two models would correctly represent the relationships between the two
causes and the effect, but the second one (after an arc has been reversed) needs more
parameters (which would also be harder to estimate and interpret). So using the proper
direction of the arcs saves work and makes the definition of the model easier. To be able
to choose the proper direction, we should think about which variables have causal
influence in which others. For instance, if X is type_of_animal and Y represents a
characteristic of an animal like presence_of_feathers, it seems clear that the type of
animal has causal influence on the presence of feathers, and therefore the more adequate
option is X  Y. However, it seems that humans find it easier to structure the information
in terms of IF-THEN rules rather than in terms of causal relationships. Because of that, it
is quite common to see the other option, perhaps influenced by the implicit construction in
our mental model of the corresponding diagnostic rule (IF it has feathers, THEN it is an
animal) instead of the causal relationship (it has feathers BECAUSE it is an animal).
In the case of student modeling, the basic overlay knowledge model usually includes
the following relationships: granularity relationships, relationships between knowledge
and evidential nodes, and prerequisite relationships. Examples of Bayesian student models
that use both modeling options (arcs in both causal and diagnostic directions) for each of
such relationships can be easily found in literature and will be discussed in the following
paragraphs. We continue using the generic notation introduced in section 1.4.1: K will
represent a knowledge node, while E will represent any event that can provide information
(evidence) about the student’s knowledge K (e.g., a question, test, exercise, task, etc.).
For modeling the relationships between the evidential nodes E and the knowledge node
K, the options are:
(o1) causal direction, i.e., K  E; or
(o2) diagnostic direction, E  K.
Option o2 could be also interpreted as a causal relationship if the variable K is
considered to be a measure of how well the student performed in a test. But usually the
goal is to evaluate the knowledge of the student and not his/her performance. In fact, the
performance can be seen as an effect of the knowledge, together with many other causes
such us the emotional state of the student, the difficulty of the evidential task being
performed, etc.
Let us see an example in the context of the adding/multiplying problem. The modeling
options are as depicted in Fig. 1.8. If we use option (o1), the nodes “Add” and “Multiply”
represent the knowledge that the student has about those two operations. In option (o2) the
nodes “Add” and “Multiply” represent an artificial constructs that measures how well the
student did in the test involving such operations.
Though this section is devoted only to qualitative model, thinking about the parameters
needed in each option can help in taking a decision. In option (o1) we need to estimate the
prior probability distributions P(Add), P(Multiply), and the conditional probability
distributions P((3+4)/Add), P((3*2+7)/Add, Multiply), and P((5*4)/Multiply). All of these
parameters have an intuitive meaning: the priors represent the “prevalence” of the two

basic operations in the population under study, i.e., how many of our target students know
them, while the conditionals represent the probability of a correct answer given that the
knowledge required has/has not been acquired. In option (o2), we would need the priors
P(3+4), P(3*2+7) and P(5*4), which can be meaningfully estimated, but also the
conditionals P(Add/(3+4), (3*2+7)), P(Multiply/(3*2+7),(3+4)), which seem harder to
interpret. For example, what is the probability that the student “performed well” in the test
with respect to the operation “Add” given that his/her answer to 3+4 was incorrect and to
3*2+7 was correct?
Add
3 + 4
Multiply
3*2 + 7
5 * 4
Add
3 + 4
Multiply
3*2 + 7
5 * 4
(o1) causal direction
(o2) diagnostic direction
Fig. 1.8. Options available for modeling relationships between knowledge and evidential nodes:
an example
Both modeling options are present in the literature. In [196] evidence is provided by
problems that the student needs to solve and knowledge nodes represent the physics rules
that the student needs to apply. The conclusion reached in the paper is that if links are
defined in the diagnostic direction it “means that the rules are conditionally independent
given the evidence, which just isn’t true (see [196] for an example). Therefore, the causal
direction seems preferable in this context. The same conclusion is drawn in [130] that
theoretically compares both options in terms of criteria such as knowledge representation,
independence relationships, reasoning process, and knowledge engineering effort
required. More recently, [119] performed an empirical study with the goal to compare the
accuracy reached with each modeling option in the diagnosis of a group of real students,
reporting a maximum accuracy of 0.508 if o2 (the diagnostic direction) is used (no better
than chance, in spite of their efforts of fine-tune the model using data) versus an average
test set accuracy of 0.776 of o1 (causal direction).
With respect to granularity relationships, options available are:
(o3) the causal direction, i.e., K1  K; or
(o4) the diagnostic direction K  K1.
The first option (o3) assumes that knowing a component indicates knowledge about the
whole, while the second option (o4) states that knowing the whole has influence in
knowing each of its components.

In our example, we can define an aggregated knowledge element “Basic arithmetic,”
which would be divided into “Add” and “Multiply.” These two modeling options are
depicted in Fig. 1.9.
Basic arithmetic
Add
Multiply
Basic arithmetic
Add
Multiply
(o3) causal direction
(o4) diagnostic direction
Fig. 1.9. Two options for modeling granularity relationships
The interpretation of both options follows. The implicit consideration under (o3) is that
if a student knows basic arithmetic, then he/she also knows how to add and how to
multiply, while (o4) assumes that a student has knowledge of basic arithmetic because
he/she knows how to add and how to multiply.
In the context of granularity relationships, examples of BNs with links defined in the
diagnostic direction (o4) are: [82; 132; 139], while [119; 122; 130; 195; 208] use links
defined in the causal direction (o3). The first theoretical comparison between both options
can be found in [90], which discusses some implications but does not explicitly
recommend any of them. The causal direction (o3) is supported by theoretical studies such
us [130], which compare both options in terms of the same criteria described above, and
empirical studies such as [39], which evaluate three different course hierarchies in the
context of adaptive testing, using quality measures such as test length and test coverage.
Another interesting alternative for the modeling of relationships between knowledge
and evidential nodes are dynamic models, which allows for changing the variables over
time. This is a desirable characteristic in the context of student modeling since the
knowledge and other relevant variables usually change continuously during a student’s
interaction with the system. In the Dynamic Bayesian Networks (DBNs) model [168], a
separate BN is constructed for each time slice. This approach is also taken in [119; 122;
130; 209], probably inspired by Reye’s work [162]. Reye presents a simple model based
on DBN, which captures the dynamic nature of the student’s knowledge. In this proposal,
for each j = 1, ..., n, ..., the following nodes and relationships between them are defined
(Fig. 1.10):
Lj = student’s state of knowledge after the j’th interaction with the system
Oj = result of the j’th interaction.

O1
L0
L1
After the first
interaction
On
Ln-1
Ln
After the n’th
interaction
On-1
...
...
...
...
Fig. 1.10. Dynamic BN for student modeling. Adapted from [162]
We can see that at time n, the probability that the student is now in the learned state
(Ln) depends on whether the student was already in the learned state at n-1 (Ln-1) and on
the outcome of the last interaction with the system (On). Reye provides formulas for a
two-phase updating of this basic model and proves that the knowledge tracing model used
in the ACT cognitive tutors [44; 45] can be viewed as a particular case of this more
general model (by adding constraints to the general model until the equivalence is shown).
Please note that in spite of the fact that this description seems to refer to only one
knowledge node, the implementations of this model will contain as many knowledge
nodes as needed, resulting in more complex networks. A later study [131] makes the
empirical comparison of a static BN with a DBN using simulated students, showing that
both models produce a very similar performance in terms of test length and accuracy,
although the static model seems to perform slightly better.
In summary, it seems that evidence available (both theoretical and empirical)
encourages the use of the causal direction. The arc should then go from variables that
cannot be observed directly and need to be estimated (knowledge in the case of student
modeling, illness or disease in the case of medical diagnosis, faulty components in the
case of trouble-shooting systems) to variables that can be observed (answers to problems,
symptoms, or problems, respectively).
With respect to prerequisite relationships, it seems clear that if A is a prerequisite of B,
knowing A has causal influence on knowing B, so the direction of the link should be A
B. However, the main difficulty of introducing this new kind of relationship into our
model is that, as reported in [31], the meaning of the relationships between the nodes
becomes somehow unclear and the specification of the parameters becomes more difficult.
To illustrate this, let us consider the following example: in a basic arithmetic course,
students are taught how to add and multiply natural numbers (N) and fractions (Q). A
basic overlay BN student model for this course (with links in the causal direction) is given
in Fig. 1.11 where prerequisite relationships are represented as a dotted line.
In this example, one of the parameters needed is P(Multiply/Mult_N, Mult_Q, Add).
But the fact that different types of relationships are mixed in the conditional distribution
makes this probability difficult to estimate. Just as an example, we would need to provide
the probability of knowing how to multiply, given that “the student knows how to
multiply natural numbers and fractions, but does not now how to add,” which is quite

improbable. As suggested in [31], a possible solution is to disregard that kind of
relationship in the model (thus making a simplification of reality, in which prerequisite
relationships do exist). Another possible solution is to make a different simplifying
assumption: instead of not including prerequisites, consider that both relationships operate
at different levels, i.e., to build a multi-layered student model as proposed by [200]. In
[31], Carmona et al present an empirical study comparing both alternatives. The study
supports a conclusion previously established in [59]: “not considering valid prerequisites
relationships does not lead to a wrong assessment of a student’s knowledge state, but it
renders the assessment less efficient in the sense that more answers than necessary have to
be collected.” So, if the test length is not an important issue, eliminating prerequisites will
not affect the accuracy of our assessment, but if length is important, a possible option
would be to use a multilayered model which can improve the performance (in the sense
that less questions will be needed to reach the same accuracy for our estimation of the
student’s knowledge).
3+2+2/3
Add N
Add Q
Add
Math
3*2
Mult N
Mult Q
Multiply
3+2
Fig. 1.11. Introducing prerequisites in the BN
Another option for considering prerequisite relationships is presented in [82] and is
based on categorizing the knowledge nodes into levels. The first level corresponds to
simple concepts, which do not require prerequisites. The successive levels require that
knowledge about some of the concepts in the previous levels be understood. In this
application, prerequisite relationships are only allowed between the root nodes of each
level (recall that this system uses the diagnostic direction (o4) for the granularity
relationships, so the root nodes are the compound concepts). In this way, the problem of
confusing different kinds of relationships in the model can be avoided.
Modeling Knowledge Beyond Overlay Models. There are also some proposals to model
student knowledge with BNs that are not based on overlay models. Alternative approaches
include constraint-based modeling and models based on misconceptions. A good example
of constraint-based modeling is the student model of CAPIT [122], a constraint-based
tutor for English capitalization and punctuation. In this model each node Li represents the

outcome of the student’s last attempt with respect to the violation or not of a certain
constraint I. That attempt can take values S (satisfied), V (violated), VFB (violated with
feedback) or NR (not relevant). When a new problem is presented, node Ni represents the
predicted outcome of this new attempt with respect to the violation of constraint i. The
authors decided to include relationships between the constraints, which caused the model
to grow more complex but was more realistic and allowed for a better prediction of
student performance. The model also includes decision nodes and their utilities, i.e.,
transforming the BN in an Influence Diagram or Decision Network (DN). This allows for
the selection of the next tutorial action by maximizing the expected utility. A similar
approach based on DN has also been presented in [138].
As for bug models, a good example is the student model for a tutoring system on
decimal numbers presented in [180]. In this case there are several types of nodes:
evidential nodes, which can be either test items of a decimal comparison test (nodes TI) or
variables that account for student behavior during the interaction with the system (nodes
B) and student-type nodes, which serve the purpose of classification of students according
to the type of misconception they have. This classification happens at two levels of
granularity: coarse and fine. The basic structure of the network is as represented in Figure
1.12.
Behavioral nodes k
Student_type nodes
Test item nodes
Coarse
class
Fine class
TI1
TI2
TI3
TIn
Behavioral
nodes 1
Bs
B2
B1
Behavioral
nodes 2
Fig. 1.12. A Bayesian student model based on misconceptions
This model features several groups of behavioral nodes. Those groups can be of
different sizes and can have different internal structures (i.e., they can be connected in
different ways). They model different instructional episodes. Please note that there is only
one group of test item nodes and student type nodes. As it can be seen, in this student

model the links follow the causal direction. It is important to note that the performance of
this student model has been evaluated by comparing the results of manual expert
diagnosis with the automatic system’s diagnosis. The comparison took into account the
results of more than 2,000 students who took the DCT test and showed an 80-90%
agreement rate (the 10% variation was due to varying the values of certain parameters).
Beyond Modeling Knowledge. The basic knowledge models presented above can be
enhanced by including information that allows for model-tracing (in addition to
knowledge-tracing). This can be achieved by adding other variables like student goals,
plans, etc. An example of the second kind of model is the Bayesian student model for the
physics domain in the tutoring system Andes [40], which divides knowledge nodes into
proposition nodes (fact and goal nodes), rule application nodes, and strategy nodes (which
allow for modeling the different methods used to solve a problem, and therefore for model
tracing). A similar approach was later adopted in [186] for the domain of medical
diagnosis. The architecture of the Bayesian student model of the Andes system is depicted
in Fig. 1.13.
Context-rule
New fact
Fact
Rule Application
Rule 1
Goal
Strategy
Fig. 1.13. General BN model for facts, rules, strategies, and goals
For example, in the physics domain, if the student knows a certain rule (such as
F=m*a) he/she will be able to use it in any context. Then, if he/she also knows certain
facts (for example, m=50kg and a=4m/s2), has a certain goal (compute the force) and a
certain strategy (way of solving the problem), then the rule will be applied to the problem
and a new fact will be inferred (in this case, that F=200N). In the medical diagnosis
example, the idea of “rule” is replaced by “medical concept” that can be applied to derive
a hypothesis (equivalent to facts in the physical domain). In both domains (physics and
medicine), the application, goal and apply nodes are used in the same way. Strategies are
modeled by the fact that students can enumerate the causal hypothesis structure in any
order. That’s why they are not explicitly introduced as nodes in the model. One difference
between these two domains is that causal relationships exist between hypotheses in the
medical application but do not exist between facts in the physics tutor.

Recently, there has been increasing interest in widening the range of user
characteristics that can be measured and used in adaptive systems. In this sense, BNs have
been used to model the student beyond cognitive features (knowledge, goals) towards a
richer variety of features, like:
• Meta-cognitive skills, for example, self-explanation [41] and exploration [29]. To
support the evaluation of meta-cognitive skills the BN student model includes nodes to
represent student’s tendency to self-explain, to explore, and also nodes to model
student’s actions that are indicative of relevant applications of such abilities. Another
recent example of modeling a meta-cognitive skill is [136], which presents a
computational framework to support learning by using analogical problem solving. To
this end, new nodes have been introduced in the Andes Bayesian student model: copy
nodes (with values correct, incorrect, no_copy) which represent if the student copied
the solution from the solved example; similarity nodes (with values trivial, non_trivial,
none) that measure the degree of similarity of fact nodes between the example solution
and the student’s solution; analogy nodes (that represent the student’s tendency to use
the minimum or maximum analogies, i.e., to try to solve the problem on their own and
consult the example only when reaching an impasse or to copy as much as possible),
and an EBLC node, which allows modeling of the user’s tendency to learn by the
Explanation Based Learning of Correctness.
• Personality traits. An example is [68], that presented a BN to model learning styles
within a web-based education system. To this end, they consider three dimensions of
Felder’s framework [64], namely perception (sensory/intuitive), processing
(active/reflective), and understanding (sequential/global) as unobservable nodes, and
several evidential nodes, such as the use of mail, forum, chats, number of examples
visited, and exam results. In this case they are using the diagnostic instead of the causal
direction for links between variables.
• Affective states. An example is the affective user model in the educational game Prime
Climb, a game designed to help children learn about number factorization while being
coached by an intelligent agent [42]. This affective user model has nodes such as: goal
nodes (that model the objective during game playing, e.g., learn math, have fun, beat
partner), action nodes (for both the player and the action), goal satisfaction nodes (to
model the degree of satisfaction that each action causes), and emotional nodes that
allow for the modeling of six of the 22 emotions described in the OCC theory of
emotions [147], namely, joy/distress (states of the node “emotion for the game”),
pride/shame (states of the node “emotion for self”) and admiration/reproach (states of
the node “emotion for agent”). In this case, links are established in the causal direction.
• Attitudes, perceptions. In [7], log-data is used to infer the student’s hidden attitudes
towards learning, learning gains, and perception of the system. To this end, the student
model contains unobservable variables that measure if the students liked the system,
learned, seriously tried to learn, wanted to finish quickly, wanted to challenge
himself/herself, was concerned with getting help, had a fear of doing the wrong action,
etc. Observable variables were: average of hints asked per problem, time between

attempts, average seconds per problem, etc. In that model also, links are pointed in the
causal direction.
Going beyond modeling the student’s knowledge can certainly provide a much better
adaptation and therefore better performance in terms of learning gains. However, as
explained in this section, the relationship between the cost of building, maintaining and
effectively using such enriched student models must be evaluated on an individual basis,
to decide whether each one is worth the gained improvement in performance of the
system.
Building the Qualitative Model versus Using Learning Techniques. In the field of
student modeling, many researchers have chosen to create structural models, with or
without the help of domain experts. Good examples are Andes [40], HYDRIVE [132],
Adele [67] and DT-Tutor [138]. However, the relationships between the variables of the
model can also be learned from studying the data. To do so, the variables of interest, X1,
X2, …, Xn for the domain being modeled, must first be identified. From that list of
variables and from a dataset composed of the sets of their values (x1, …, xn), a structural
learning algorithm infers the relationships between the variables, according to a desired
confidence level. A key point of this approach is to choose a good value for the
confidence level. If too low a value is selected, low data-evidence relationships will be
inferred. If the value is too high, some important relationships can be missed. A possible
solution is having domain experts fine-tune the inferred model.
There are few examples in which the structure has been discovered from studying the
data, but the interested reader can find an excellent one in [122], which describe the
techniques used and the strengths and weaknesses of this approach.
1.4.3 Development of the Quantitative Model
Once the qualitative part of the model has been defined or learned, the next step is to
define the parameters, which in this case are the prior probability distributions of the root
nodes and the conditional probability distributions for the rest of the nodes. This task has
commonly been cited as one of the main difficulties when building a BN model. The
options available are: a) knowledge engineering, i.e., having experts specify the
probabilities; b) using pre-existent models to specify part of the probability distributions
needed (canonical models, theoretical models, etc.); and c) learning the parameters from
available data (cases). A combination of these alternatives can also be used.
Examples of systems in which the parameters have been estimated by experts include
those mentioned in subsection 1.4.2 as examples of systems in which experts provided
also the qualitative model. However, there are other systems where the structure has been
specified or constrained by experts and the parameters are partially or totally derived from
data [201], or adjusted to theoretical models, for example, models inspired by Item
Response Theory [130] and [90]. Some approaches to simplify parameter specification in

the context of student modeling have also been proposed [129]. Finally, an example in
which both the structure and parameters have been discovered from data is [122].
Also, some modeling tricks can help to relieve the burden of specifying the numbers.
An example of such a trick is grouping the parent nodes into related causal categories,
which reduces the number of parameters needed. For a more comprehensive literature
guide to tricks and techniques, see [60].
1.4.4 Building Student Models by Combining the Domain Expert’s Knowledge and
Learning Methods
More recently, researchers have begun to use mixed approaches, in which parts of the
model are defined by experts while others are learned from data. Some examples are:
• In the student model for the medical domain by Suebnukarn et al [186], the structure
has been specified by experts and then the parameters are mined from the data obtained
in transcripts of problem-based learning sessions.
• Stacey et al construct their Bayesian student model based on misconceptions through a
combination of elicitation from the domain experts and automated methods [180].
• When developing the knowledge student model for Prime Climb, [119] compares the
performance of two different structures (one developed following teacher’s
suggestions, the other one inspired by causality) and for both of them the parameters
are revealed in the data.
• Arroyo and Woolf define the structure of the network, taking into account the
knowledge gained during a correlation analysis between variables, and then look for
conditional probabilities in the data to evaluate the accuracy of the model by cross
validation [7].
• In [29] and [42], Conati and colleagues use an iterative design process: the initial
model is developed using the researcher’s intuition or the expert’s opinions, and then
data from real users is collected and used to refine both the parameters and the
structure, either by adding new nodes (such as “general exploration” or nodes that
describe attitudes towards the intelligent agent, such as “wanting help”) or by
modifying the direction of the links (for example, between knowledge and exploration
nodes or game events and goals).
1.5 Conclusions
A common feature of various adaptive Web systems is the application of user models
(also known as profiles) to adapt the systems’ behavior to individual users. User models
represent the information about users that is essential to support the adaptation
functionality of the systems. Adaptive Web systems have investigated a range of
approaches to user modeling, exploring how to organize the storage for user information,

how to populate it with user data, and how to maintain the current state of the user. The
majority of modern adaptive Web systems use feature-based approach to represent and
model information about the users. The competing stereotype-based approach, once
popular in the pre-Web area of adaptive interfaces, has lost dominance but is still applied,
especially in combination with the feature-based approaches.
The most popular features modeled and used by adaptive Web systems are user
knowledge, interests, goals, background, individual traits, and context of work. Each
individual adaptive system typically uses a subset of this list, as determined by the class of
adaptive systems it belongs to and the adaptation needs of this class (Fig. 1.14). Web-
based adaptive educational systems (AES) rely mostly on user knowledge and learning
goals capitalizing on the modeling and representation techniques established in the field of
Intelligent Tutoring Systems (ITS). Adaptive information systems and Web
recommenders focus on modeling the user’s interests and extend modeling approaches
originally developed for adaptive information retrieval systems. Meanwhile, adaptive
hypermedia systems attempt to represent and employ an even wider range of user features.
In addition to user knowledge and interests, these systems frequently model user goals
(following approaches developed in the field of adaptive interfaces), individual traits, and
the context of user’s work.
Fig. 1.14. User features typically modeled by different classes of adaptive Web systems
Overlay user modeling is currently the leading user modeling approach in AES and
AHS. This approach was originally developed in the field of ITS to model user knowledge
as an overlay of a concept-level domain model. Currently, the overlay approach has
grown to include the modeling of knowledge, interests, goals, and some other features. In
the area of adaptive information systems the concept-overlay approach to modeling user
interests is now competing with the more traditional keyword-level profiling. This is one
example of the convergence that has begun to blur the boundaries between different
classes of adaptive Web systems.
One of the main advantages of using a formal method for student modeling is its
robustness. Once this model behaves in a stable and theoretically-correct fashion, the

evaluation of a system can be focused on other components (such as quality of the
learning material, learning strategies used, or adaptation capabilities). The most
commonly used formal reasoning technique for student modeling is currently the BN
paradigm. Applications of other approaches (like Fuzzy Logic) are less frequent,. A very
attractive potential use of BNs in the context of adaptive web-based applications would be
to employ learning algorithms that would further shape on-the-fly improvements within
the model itself. The future adaptation learning algorithm would then be able to process
each user’s interactive behavior information and simultaneously update the structure of
the model. We hope to see this potential use developed in the very near future.
References
1. Anderson, J.R., Corbett, A.T., Ko edinger, K.R., Pelletier, R.: Cognitive tutors: Lessons learned.
The Journal of the Learning Sciences 4, 2 (1995) 167-207
2. Anderson, J.R., Lebiere, C.: Atomic Components of Thought. Lawrence Erlbaum Associates,
Hillsdale, NJ (1998)
3. Ardissono, L., Console, L., Torre, I.: An adaptive system for the personalised access to news. AI
Communications 14 (2001) 129-147
4. Ardissono, L., Goy, A.: Tailoring the interaction with users in Web stores. User Modeling and
User Adapted Interaction 10, 4 (2000) 251-303
5. Ardissono, L., Goy, A., Meo, R., Petrone, G., Console, L., Lesmo, L., Simone, C., Torasso, P.: A
Configurable System for the Construction of Adaptive Virtual Stores. World Wide Web 2, 3
(1999) 143-159
6. Aroyo, L., Dicheva, D.A.: Concept-based approach to support learning in a Web-based support
environment. In: Moore, J.D., Redfield, C.L., Johnson, W.L. (eds.) Proc. of AI-ED'2001. IOS
Press (2001) 1-12
7. Arroyo, I., Woolf, B.P.: Inferring learning and attitudes from a Bayesian Network of log file
data. In: Looi, C.-K., McCalla, G., Bredeweg, B., Breuker, J. (eds.) Proc. of 12th International
Conference on Artificial Intelligence in Education, AIED'2005. IOS Press (2005) 33-40
8. Beaumont, I.: User modeling in the interactive anatomy tutoring system ANATOM-TUTOR.
User Modeling and User-Adapted Interaction 4, 1 (1994) 21-45
9. Beck, J.: Engagement tracing: using response times to model student disengagement. In: Looi,
C.-K., McCalla, G., Bredeweg, B., Breuker, J. (eds.) Proc. of 12th International Conference on
Artificial Intelligence in Education, AIED'2005. IOS Press (2005) 88-95
10.Benyon, D., Murray, D.: Experience with adaptive interfaces. Th e Computer Journal 31, 5 (1988)
465-473
11.Berti, S., Mori, G., Paternò, F., Santoro, C.: An environment for designing and developing multi-
platform interactive applications. In: Ardissono, L., Goy, A. (eds.) Proc. of HCITALY'2003.
University of Turin (2003) 7-16
12.Billsus, D., Pazzani, M.J.: A learning agent for wireless news access. In: Lieberman, H. (ed.)
Proc. of 2000 International Conference on Intelligent User Interfaces. ACM Press (2000) 94-97
13.Bloom, B.S.: Taxonomy of Educational Objectives, Handbook I: The Cognitive Domain. David
McKay Co Inc., New York (1956)

14.Boyle, C., Encarnacion, A.O.: MetaDoc: an adaptive hypertext reading system. User Modeling
and User-Adapted Interaction 4, 1 (1994) 1-19
15.Brady, A., Conlan, O., Wade, V.: Dynamic Composition and Personalization of PDA-based
eLearning – Personalized mLearning. In: Nall, J., Robson, R. (eds.) Proc. of World Conference
on E-Learning, E-Learn 2004. AACE (2004) 234-242
16.Brailsford, T.J., Stewart, C.D., Zakaria, M.R., Moore, A.: Autonavigation, links, and narrative in
an adaptive Web-based integrated learning environment. In: Proc. of The 11th International
World Wide Web Conference. (2002)
17.Brajnik, G., Guida, G., Tasso, C.: User modeling in intelligent information retrieval. Information
Processing and Management 23, 4 (1987) 305-320
18.Brusilovsky, P.: A framework for intelligent knowledge sequencing and task sequencing. In:
Frasson, C., Gauthier, G., McCalla, G. (eds.) Proc. of Second International Conference on
Intelligent Tutoring Systems, ITS'92. Springer-Verlag (1992) 499-506
19.Brusilovsky, P.: Adaptive hypermedia, an attempt to analyze and generalize. In: Brusilovsky, P.,
Kommers, P., Streitz, N. (eds.): Multimedia, Hypermedia, and Virtual Reality. Lecture Notes in
Computer Science, Vol. 1077. Springer-Verlag, Berlin (1996) 288-304
20.Brusilovsky, P.: Developing Adaptive Educational Hypermedia Systems: From Design Models
to Authoring Tools. In: Murray, T., Blessing, S., Ainsworth, S. (eds.): Authoring Tools for
Advanced Technology Learning Environments: Toward cost-effective adaptive, interactive, and
intelligent educational software. Dordrecht, Kluwer (2003) 377-409
21.Brusilovsky, P.: Adaptive navigation support. In: Brusilovsky, P., Kobsa, A., Neidl, W. (eds.):
The Adaptive Web: Methods and Strategies of Web Personalization. Lecture Notes in Computer
Science, Vol. 4321. Springer-Verlag, Berlin Heidelberg New York (2007) this volume
22.Brusilovsky, P., Anderson, J.: ACT-R electronic bookshelf: An adaptive system for learning
cognitive psychology on the Web. In: Maurer, H., Olson, R.G. (eds.) Proc. of WebNet'98, World
Conference of the WWW, Internet, and Intranet. AACE (1998) 92-97
23.Brusilovsky, P., Cooper, D.W.: Domain, Task, and User Models for an Adaptive Hypermedia
Performance Support System. In: Gil, Y., Leake, D.B. (eds.) Proc. of 2002 International
Conference on Intelligent User Interfaces. ACM Press (2002) 23-30
24.Brusilovsky, P., Eklund, J., Schwarz, E.: Web-based education for all: A tool for developing
adaptive courseware. In: Ashman, H., Thistewaite, P. (eds.) Proc. of Seventh International World
Wide Web Conference. Vol. 30. Elsevier Science B. V. (1998) 291-300
25.Brusilovsky, P., Henze, N.: Open corpus adaptive educational hypermedia. In: Brusilovsky, P.,
Kobsa, A., Neidl, W. (eds.): The Adaptive Web: Methods and Strategies of Web Personalization.
Lecture Notes in Computer Science, Vol. 4321. Springer-Verlag, Berlin Heidelberg New York
(2007) this volume
26.Brusilovsky, P., Schwarz, E., Weber, G.: ELM-ART: An intelligent tutoring system on World
Wide Web. In: Frasson, C., Gauthier, G., Lesgold, A. (eds.) Proc. of Third International
Conference on Intelligent Tutoring Systems, ITS-96. Lecture Notes in Computer Science, Vol.
1086. Springer Verlag (1996) 261-269
27.Brusilovsky, P., Sosnovsky, S., Yudelson, M.: Ontology-based framework for user model
interoperability in distributed learning environments. In: Richards, G. (ed.) Proc. of World
Conference on E-Learning, E-Learn 2005. AACE (2005) 2851-2855
28.Bunt, A., Carenini, G., Conati, C.: Adaptive content presentation for the Web. In: Brusilovsky,
P., Kobsa, A., Neidl, W. (eds.): The Adaptive Web: Methods and Strategies of Web
Personalization. Lecture Notes in Computer Science, Vol. 4321. Springer-Verlag, Berlin
Heidelberg New York (2007) this volume

29.Bunt, A., Conati, C.: Probabilistic Student Modelling to Improve Exploratory Behaviour. User
Modeling and User-Adapted Interaction 13, 3 (2003) 269-309
30.Carberry, S., de Rosis, F. (eds.): Proceedings of Wo rkshop on Adapting the Interaction Style to
Affective Factors hold in conjunction with User Modeling 2005, July 25, 2005. Edinburgh, UK
(2005) available online at http://www.di.uniba.it/intint/UM05/WS-UM05.html
31.Carmona, C., Millán, E., Perez de la Cruz, J.-L., Trella, M., Conejo, R.: Introducing Prerequisite
Relations in a Multi-layered Bayesian Student Model. In: Proc. of 10th International Conference
UM'2005. Vol. 3538. Springer-Verlag (2005) 347-356
32.Carro, R.M., Pulido, E., Rodríguez, P.: Dynamic generation of adaptive Internet-based courses.
Journal of Network and Computer Applications 22, 4 (1999) 249-257
33.Carver, C.A., Howard, R.A., Lavelle, E.: Enhancing student learning by incorporating student
learning styles into adaptive hypermedia. In: Proc. of ED-MEDIA'96 - World Conference on
Educational Multimedia and Hypermedia. AACE (1996) 118-123
34.Cawsey, A., Grasso, F., Paris, C.: Adaptive information for consumers of healthcare. In:
Brusilovsky, P., Kobsa, A., Neidl, W. (eds.): The Adaptive Web: Methods and Strategies of Web
Personalization. Lecture Notes in Computer Science, Vol. 4321. Springer-Verlag, Berlin
Heidelberg New York (2007) this volume
35.Chen, S.Y., Macredie, R.D.: Cognitive styles and hypermedia navigation: Development of a
learning model. Journal of the American Society for Information Science and Technology 53, 1
(2002) 3-15
36.Cheverst, K., Davies, N., Mitchell, K., Smith, P.: Providing tailored (context-aware) information
to city visitors. In: Brusilovsky, P., Stock, O., Strapparava, C. (eds.) Proc. of Adaptive
Hypermedia and Adaptive Web-based Systems. AH'2000. Lecture Notes in Computer Science,
Vol. 1892. Springer-Verlag (2000) 73-85
37.Chin, D., Kobsa, A., Wahlster, W.: Modelling what the User Knows in UC. In: Loveland, D.W.
(ed.) User Models in Dialog Systems. Symbolic Computation Series, Springer Verlag, Berlin
(1989) 74-107
38.Cocea, M., Weibelzahl, S.: Can log files analysis estimate learners’ level of motivation? In:
Herder, E., Heckmann, D. (eds.) Proc. of 14th Workshop on Adaptivity and User Modeling in
Interactive Systems, ABIS 2006. University of Hildesheim (2006) 32-35
39.Collins, J.A., Greer, J.E., Huang, S.H.: Adaptive Assessment Using Granularity Hierarchies and
Bayesian Nets. In: Frasson, C., Gauthier, G., Lesgold, A. (eds.) Proc. of 3rd International
Conference on Intelligent Tutoring Systems, ITS'96. Lecture Notes in Computer Science, Vol.
1086. Springer-Verlag (1996) 569-577
40.Conati, C., Gertner, A., VanLehn, K.: Using Bayesian Networks to Manage Uncertainty in
Student Modeling. User Modeling and User-Adapted Interaction 12, 4 (2002) 371-417
41.Conati, C., Larkin, J., VanLehn, K.: A Computer Framework to Support Self-explanation. In: du
Bolay, B., Mizoguchi, R. (eds.) Proc. of 8th World Conference on Artificial Intelligence in
Education AIED'97. Knowledge and Media in Learning Systems, IOS Press (1997) 279-286
42.Conati, C., Maclaren, H.: Data-Driven Refinement of a Probabilistic Model of User Affect. In:
Proc. of 10th International Conference UM'05. Lecture Notes in Computer Science, Vol. 3538.
Springer-Verlag (2005) 40-49
43.Conlan, O., O'Keeffe, I., Tallon, S.: Combining adaptive hypermedia techniques and ontology
reasoning to produce dynamic personalized news services. In: Wade, V., Ashman, H., Smyth, B.
(eds.) Proc. o f 4th International Conference on Adaptive Hypermedia and Adaptive Web-Based
Systems (AH'2006). Lecture Notes in Computer Science, Vol. 4018. Springer Verlag (2006) 81-
90

44.Corbett, A., Anderson, J.: Student Modelling and Mastery Learning in a Computer-based
Programming Tutor. In: Frasson, C., Gauthier, G., McCalla, G.I. (eds.) Proc. of 2nd International
Conference on Intelligent Tutoring Systems, ITS'92. Lecture Notes in Computer Science, Vol.
608. Springer-Verlag (1992) 413-420
45.Corbett, A., Anderson, J.R., O'Brien, A.T.: Student Modeling in the ACT Programming Tutor.
In: Cognitively Diagnostic Assessment. Erlbaum, Hillsdale, NJ (1995) 19-41
46.Cotter, P., Smyth, B.: WAP-ing the Web: Content personalization for WAP-enabled devices. In:
Brusilovsky, P., Stock, O., Strapparava, C. (eds.) Proc. of Adaptive Hypermedia and Adaptive
Web-based systens. Lecture Notes in Computer Science, Vol. 1892. Springer-Verlag (2000) 98-
108
47.Dagger, D., Conlan, O., Wade, V.P.: An architecture for candidacy in adaptive eLearning
systems to facilitate the reuse of learning Resources. In: Rossett, A. (ed.) Proc. of World
Conference on E-Learning, E-Learn 2003. AACE (2003) 49-56
48.De Bra, P., Aerts, A., Rousseau, B.: Concept Relationship Types for AHA! 2.0. In: Driscoll, M.,
Reeves, T.C. (eds.) Proc. of World Conference on E-Learning, E-Learn 2002. AACE (2002)
1386-1389
49.De Bra, P., Aerts, A., Smits, D., Stash, N.: AHA! Version 2.0: More Adaptation Flexibility for
Authors. In: Driscoll, M., Reeves, T.C. (eds.) Proc. of World Conference on E-Learning, E-Learn
2002. AACE (2002) 240-246
50.De Bra, P., Calvi, L.: AHA! An open Adaptive Hypermedia Architecture. The New Review of
Hypermedia and Multimedia 4 (1998) 115-139
51.De Bra, P., Ruiter, J.-P.: AHA! Adaptive hypermedia for all. In: Fowler, W., Hasebrook, J. (eds.)
Proc. of WebNet'2001, World Conference of the WWW and Internet. AACE (2001) 262-268
52.De Bra, P.M.E.: Teaching Hypertext and Hypermedia through the Web. Journal of Universal
Computer Science 2, 12 (1996) 797-804
53.de Rosis, F., De Carolis, B., Pizzutilo, S.: U ser tailored hypermedia explanations. In:
Brusilovsky, P., Beaumont, I. (eds.) Proc. of Workshop Adaptive Hypertext and Hypermedia at
Fourth International Conference on User Modeling. (1994)
http://wwwis.win.tue.nl/ah94/deRosis.html
54.Díaz, A., Gervás, P.: Personalisation in news delivery systems: Item summarization and multi-
tier item selection using relevance feedback. Web Intelligence and Agent Systems 3, 3 (2005)
135-154
55.Dichev, C., Dicheva, D., Aroyo, L.: Using Topic Maps for Web-based Education. Advanced
Technology for Learning 1, 1 (2004) 1-7
56.Dolog, P., Henze, N., Nejdl, W., Sintek, M.: Personalization in distributed e-learning
environments. In: Proc. of The Thirteenth International World Wide Web Conference, WWW
2004 (Alternate track papers and posters). ACM Press (2004) 161-169
57.Dolog, P., Nejdl, W.: Semantic Web Technologies for the Adaptive Web. In: Brusilovsky, P.,
Kobsa, A., Neidl, W. (eds.): The Adaptive Web: Methods and Strategies of Web Personalization.
Lecture Notes in Computer Science, Vol. 4321. Springer-Verlag, Berlin Heidelberg New York
(2007) this volume
58.Dolog, P., Schäfer, M.: Learner Modeling on the Semantic Web. In: Proc. of PerSWeb’05,
Workshop on Personalization on the Semantic Web at 10th International User Modeling
Conference. (2005) http://www.win.tue.nl/persweb/Camera-ready/6-Dolog-full.pdf
59.Dowling, C.E., Hockemeyer, C., Ludwig, A.H.: Adaptive Asessment and Training Using the
Neighbourhood of Knowledge States. In: Frasson, C., Gauthier, G., Lesgold, A. (eds.) Proc. of

3rd International Conference on Intelligent Tutoring Systems, ITS'96. Lecture Notes in
Computer Science, Vol. 1086. Springer-Verlag (1996) 578-585
60.Druzdzel, M., van der Gaag, L.C.: Building Probabilistic Networks: Where Do the Numbers
Come From? - Guest editors' introduction. IEEE Transactions on Knowledge and Data
Engineering 12, 4 (2000) 481-486
61.Dufresne, A., Turcotte, S.: Cognitive style and its implications for navigation strategies. In: du
Boulay, B., Mizoguchi, R. (eds.) Proc. of AI-ED'97, 8th World Conference on Artificial
Intelligence in Education. IOS (1997) 287-293
62.Eisenstein, J., Vanderdonckt, J., Puerta, A.: Applying model-based techniques to the
development of UIs for mobile computers. In: Proc. of 6th International Conference on
Intelligent User Interfaces. ACM Press (2001) 69-76
63.Encarnação, L.M.: Multi-level user support through adaptive hypermedia: A highly application-
independent help component. In: Moore, J., Edmonds, E., Puerta, A. (eds.) Proc. of 1997
International Conference on Intelligent User Interfaces. ACM (1997) 187-194
64.Felder, R.: Learning and teaching styles. Journal of Engineering Education 78, 7 (1988) 674-681
65.Fink, J., Kobsa, A., Nill, A.: Adaptable and adaptive information provision for all users,
including disabled and elderly people. The New Review of Hypermedia and Multimedia 4 (1998)
163-188
66.Fischer, G.: User modeling in human-computer interaction. User Modeling and User Adapted
Interaction 11, 1-2 (2001) 65-86
67.Ganeshan, R., Johnson, W., Shaw, E., Wood, B.P.: Tutoring Diagnostic Problem Solving. In:
Proc. of 7th International Conference on Intelligent Tutoring Systems, ITS'2004. Vol. 1839.
Springer-Verlag (2000) 33-42
68.García, F., Amandi, A., Schiaffinoa, S., Campoa, M.: Evaluating Bayesian networks' precision
for detecting students' learning styles. Computers & Education (2006) In press
69.Garlatti, S., Iksal, S.: Context filtering and spacial filtering in an adaptive information system. In:
Brusilovsky, P., Stock, O., Strapparava, C. (eds.) Proc. of Adaptive Hypermedia and Adaptive
Web-based systens. Lecture Notes in Computer Science, Vol. 1892. Springer-Verlag (2000) 315-
318
70.Garlatti, S., Iksal, S., Kervella, P.: Adaptive on-line information system by means of a task
model and spatial views. In: Brusilovsky, P., Bra, P.D. (eds.) Proc. of Second Workshop on
Adaptive Systems and User Modeling on the World Wide Web. (1999) 59-66, also available at
http://wwwis.win.tue.nl/asum99/garlatti/garlatti.html
71.Gates, K.F., Lawhead, P.B., Wilkins, D.E.: Toward an adaptive WWW: a case study in
customized hypermedia. New Review of Multimedia and Hypermedia 4 (1998) 89-113
72.Gauch, S., Speretta, M., Chandramouli, A., Micarelli, A.: User profiles for personalized
information access. In: Brusilovsky, P., Kobsa, A., Neidl, W. (eds.): The Adaptive Web:
Methods and Strategies of Web Personalization. Lecture Notes in Computer Science, Vol. 4321.
Springer-Verlag, Berlin Heidelberg New York (2007) this volume
73.Gilbert, J.E., Han, C.Y.: Arthur: Ad apting Instruction to Accommodate Learning Style. In: Bra,
P.D., Leggett, J. (eds.) Proc. of WebNet'99, World Conference of the WWW and Internet. AACE
(1999) 433-438
74.Goldstein, I.P.: The genetic graph: a representation for the evolution of procedural knowledge.
In: Sleeman, D.H., Brown, J.S. (eds.): Intelligent tutoring systems. Academic press, London
(1982) 51-77

75.Gonschorek, M., Herzog, C.: Using hypertext for an adaptive helpsystem in an intelligent
tutoring system. In: Greer, J. (ed.) Proc. of AI-ED'95, 7th World Conference on Artificial
Intelligence in Education. AACE (1995) 274-281
76.Goodman, B.A., Litman, D.J.: On the interaction between plan recognition and intelligent
interfaces. User Modeling and User-Adapted Interaction 2, 1 (1992) 83-115
77.Goren-Bar, D., Graziola, I., Pianesi, F., Zancanaro, M.: The influence of personality factors on
visitor attitudes towards adaptivity dimensions for mobile museum guides. User Modeling and
User Adapted Interaction 16, 1 (2005) 31-62
78.Grunst, G.: Adaptive hypermedia for support systems. In: Schneider-Hufschmidt, M., Kühme, T.,
Malinowski, U. (eds.): Adaptive user interfaces: Principles and practice. North-Holland,
Amsterdam (1993) 269-283
79.Heckerman, D.: A Tutorial on Learning with Bayesian Networks, Technical Report No. MSR-
TR-95-06, Microsoft Research Advanced Technology Division (1995)
80.Heckmann, D., Schwartz, T., Brandherm, B., Schmitz, M., von Wilamowitz-Moellendorff, M.:
Gumo - The General User Model Ontology. In: Ardissono, L., Brna, P., Mitrovic, A. (eds.) Proc.
of 10th International User Modeling Conference. Lecture Notes in Artificial Intelligence, Vol.
3538. Springer Verlag (2005) 428-432
81.Henze, N., Naceur, K., Nejdl, W., Wolpers, M.: Adaptive hyperbooks for constructivist teaching.
Künstliche Intelligenz, 4 (1999) 26-31
82.Henze, N., Nejdl, W.: Student modeling for KBS Hyperbook system using Bayesian networks,
Technical report, University of Hannover (1999) available online at http://www.kbs.uni-
hannover.de/paper/99/adaptivity.html
83.Henze, N., Nejdl, W.: Adaptation in open corpus hypermedia. International Journal of Artificial
Intelligence in Education 12, 4 (2001) 325-350
84.Hirashima, T., Hachiya, K., Kashihara, A., Toyoda, J.i.: Information filtering using user's context
on browsing in hypertext. User Modeling and User Adapted Interaction 7, 4 (1997) 239-256
85.Hirashima, T., Matsuda, N., Nomoto, T., Toyoda, J.i.: Context-sensitive filtering for browing in
hypertext. In: Proc. of International Conference on Intelligent User Interfaces, IUI'98. ACM
Press (1998) 21-28
86.Hockemeyer, C., Held, T., Albert, D .: RATH - A relational adaptive tutoring hypertext WWW-
environment based on knowledge space theory. In: Alvegård, C. (ed.) Proc. of CALISCE'98, 4th
International conference on Computer Aided Learning and Instruction in Science and
Engineering. (1998) 417-423
87.Hollink, V., Someren, M.v., Hage, S.t.: Discovering stages in web navigation. In: Ardissono, L.,
Brna, P., Mitrovic, A. (eds.) Proc. of 10th International User Modeling Conference. Lecture
Notes in Artificial Intelligence, Vol. 3538. Springer Verlag (2005) 473-482
88.Höök, K., Karlgren, J., Wærn, A., Dahlbäck, N., Jansson, C.G., Karlgren, K., Lemaire, B.: A
glass box approach to adaptive hypermedia. User Modeling and User-Adapted Interaction 6, 2-3
(1996) 157-184
89.Horvitz, E.J., Breese, J.S., Henrion, M.: Decision Theory in Expert Systems and Artificial
Intelligence. International Journal of Approximate Reasoning 2 (1988) 247-302
90.Jameson, A.: Numerical Uncertainty Management in User and Student Modeling: An Overview
of Systems and Issues. User Modeling and User-Adapted Interaction 5 (1996) 193-251
91.Jameson, A.: Mod eling both the context and the user. Personal Technologies 5, 1 (2001) 29-33
92.Jameson, A., Smyth, B.: Recommendation to groups In: Brusilovsky, P., Kobsa, A., Neidl, W.
(eds.): The Adaptive Web: Methods and Strategies of Web Personalization. Lecture Notes in
Computer Science, Vol. 4321. Springer-Verlag, Berlin Heidelberg New York (2007) this volume

93.Jantke, K.P., Memmel, M., Rostanin, O., Rudolf, B.: Media and service integration for
professional e-learning. In: Nall, J., Robson, R. (eds.) Proc. of World Conference on E-Learning,
E-Learn 2004. AACE (2004) 725-732
94.Jin, X., Zhou, Y., Mobasher, B.: Task-Oriented Web User Modeling for Recommendation. In:
Ardissono, L., Brna, P., Mitrovic, A. (eds.) Proc. of 10th International User Modeling
Conference. Lecture Notes in Artificial Intelligence, Vol. 3538. Springer Verlag (2005) 109-118
95.Joerding, T.: A temporary user modeling approach for adaptive shopping on the Web. In:
Brusilovsky, P., De Bra, P. (eds.) Proc. of Second Workshop on Adaptive Systems and User
Modeling on the World Wide Web. (1999) 75-79, also available at
http://wwwis.win.tue.nl/asum99/joerding/joerding.html
96.Jokela, S., Turnpeinen, M., Kurki, T., Savia, E., Sulonen, R.: The Role of Structured Content in a
Personalised News Service. In: Proc. of 34th Hawaii International Conference on System
Sciences. (2001) 1-10
97.Kaplan, C., Fenwick, J., Chen, J.: Adaptive hypertext navigation based on user goals and context.
User Modeling and User-Adapted Interaction 3, 3 (1993) 193-220
98.Katz, S., Lesgold, A., Eggan, G., Gordin, M.: Modelling the student in SHERLOCK II. In: Greer,
J.E., McCalla, G. (eds.): Student Modelling: Th e Key to Individualized Knowledge-Based
Instruction. Series F: Computer and Systems Sciences. NATO ASI Series, Springer Verlag,
Berlin Heidelberg (1994) 99-125
99.Kavcic, A.: Fuzzy user modeling for adaptation in educational hypermedia. IEEE Transactions
on Systems, Man, and Cybernetics 34, 4 (2004) 439-449
100. Kawai, K., Mizoguchi, R., Kakusho, O., Toyoda, J.: A framework for ICAI systems based
on inductive inference and logic programming. New Generation Computing 5 (1987) 115-129
101. Kay, J.: Lies, damned lies and stereotypes: pragmatic approximations of users. In: Kobsa,
A., Litman, D. (eds.) Proc. of Fourth International Conference on User Modeling. MITRE (1994)
175-184
102. Kay, J., Kummerfeld, R.J.: An individualised course for the C p rogramming language. In:
Proc. of Second International WWW Conference. (1994) http://www.cs.usyd.edu.au/~bob/kay-
kummerfeld.html
103. Kay, J., Lum, A.: Ontologies for Scrutable Learner Modeling in Adaptive E -Learning. In:
Aroyo, L., Tasso, C. (eds.) Proc. of Workshop on Application of Semantic Web Technologies for
Adaptive Educational Hypermedia at the Third International Conference on Adaptive
Hypermedia and Adaptive Web-Based Systems (AH'2004). Technische University Eindhoven
(2004) 292-301
104. Kim, K.S., Allen, B.: Cognitive and Task Influences on Web Searching Behavior. Journal
of the American Society for Information Science and Technology 53, 2 (2002) 109-119
105. Klyne, G., Reynolds, F., Woodrow, C., Ohto, H., Johan Hjelm, Butler, M.H., Tran, L.:
Composite Capability/Preference Profiles (CC/PP): Structure and Vocabularies 1.0. W3C
Recommendation 15 January 2004. (2004) http://www.w3.org/TR/CCPP-struct-vocab/
106. Koedinger, K.R., Anderson, J.R., Hadley, W.H., Mark, M.A.: Intelligent tutoring goes to
school in the big city. International Journal of Artificial Intelligence in Education 8 (1997) 30-43
107. Korfhage, R.R.: Information storage and retrieval. Wiley Computer Publishing, N.Y.
(1997)
108. Kosba, E., Dimitrova, V., Boyle, R.: Using Fuzzy Techniques to Model Students in Web-
Based Learning Environments. International Journal of Artificial Intelligence Tools 13, 2 (2004)
279-297

109. Krüger, A., Baus, J., Heckmann, D., Kruppa, M., Wasinger, R.: Adaptive mobile guides.
In: Brusilovsky, P., Kobsa, A., Neidl, W. (eds.): The Adaptive Web: Methods and Strategies o f
Web Personalization. Lecture Notes in Computer Science, Vol. 4321. Springer-Verlag, Berlin
Heidelberg New York (2007) this volume
110. Kumar, A.N.: A Scalable Solution for Adaptive Problem Sequencing and its Evaluation.
In: Wade, V., Ashman, H., Smyth, B. (eds.) Proc. of 4th International Conference on Adaptive
Hypermedia and Adaptive Web-Based Systems (AH'2006). Lecture Notes in Computer Science,
Vol. 4018. Springer Verlag (2006) 161-171
111. Laroussi, M., Benahmed, M.: Providing an adaptive learning through the Web case of
CAMELEON: Computer Aided MEdium for LEarning on Networks. In: Alvegård, C. (ed.) Proc.
of CALISCE'98, 4th International conference on Computer Aided Learning and Instruction in
Science and Engineering. (1998) 411-416
112. Lee, K.B., Grice, R.A.: An Adaptive Viewing Application for the Web on Personal
Digital Assistants. In: Proc. of ACM SIGDOC'03. IEEE (2003) 125-132
113. Liu, Y., Ginther, D.: Cognitive styles and distance education. Online Journal of Distance
Learning Administration 2, 3 (1999) http://www.westga.edu/~distance/liu23.html
114. Locke, J.: Microsoft Bayesian Networks. Basics of Knowledge Engineering. Microsoft
Support Technology (1999) http://freelock.com/files/KE.pdf
115. Lopez, J.F., Szekely, P.: Web page adaptation for universal access. In: Stephanidis, C.
(ed.) Proc. of 1st International Conference on Universal Access in Human-Computer Interaction.
Lawrence Erlbaum Associates (2001) 690-694
116. López, J.M., Millán, E., Pérez-de-la-Cruz, J.-L., Triguero, F.: ILESA: a Web-based
Intelligent Learning Environment for the Simplex Algorithm. In: Alvegård, C. (ed.) Proc. of
CALISCE'98, 4th International conference on Computer Aided Learning and Instruction in
Science and Engineering. (1998) 399-406
117. Lundgren-Cayrol, K., Paquette, G., Miara, A., Bergeron, F., Rivard, J., Rosca, I.: Explor@
Advisory Agent: Tracing the Student’s Trail. In: Fowler, W., Hasebrook, J. (eds.) Proc. of
WebNet'2001, World Conference of the WWW and Internet. AACE (2001) 802-808
118. Magnini, B., Strapparava, C.: Improving user modeling with content-based techniques. In:
Bauer, M., Gmytrasiewicz, P.J., Vassileva, J. (eds.) Proc. of 8th International Conference on
User Modeling, UM 2001. Lecture Notes on Artificial Intelligence, Vol. 2109. Springer-Verlag
(2001) 74-83
119. Manske, M., Conati, C.: Modelling Learning in an Educational Game. In: Proc. of 12th
World Conference of Artificial Intelligence and Education AIED'05. IOS Press (2005) 411-419
120. Masthoff, J.: Towards an authoring coach for adaptive Web-based instruction. In: De Bra,
P., Brusilovsky, P., Conejo, R. (eds.) Proc. of Second International Conference on Adaptive
Hypermedia and Adaptive Web-Based Systems (AH'2002). Lecture Notes in Computer Science,
Vol. 2347. (2002) 415-418
121. Mathé, N., Chen, J.: User-centered indexing for adaptive information access. User
Modeling and User-Adapted Interaction 6, 2-3 (1996) 225-261
122. Mayo, M., Mitrovic, A.: Optimising ITS behaviour with Bayesian networks and decision
theory. International Journal of Artificial Intelligence in Education 12 (2001) 124-153
123. McArthur, D., Stasz, C., Hotta, J., Peter, O., Burdorf, C.: Skill-oriented task sequencing in
an intelligent tutor for basic algebra. Instructional Science 17, 4 (1988) 281-307
124. McCalla, G., Bunt, R.B., Harms, J.J.: The design of the SCENT automated advisor.
Computational Intelligence 2, 2 (1986) 76-91

125. Micarelli, A., Gasparetti, F., Sciarrone, F., Gauch, S.: Personalized search on the World
Wide Web. In: Brusilovsky, P., Kobsa, A., Neidl, W. (eds.): The Adaptive Web: Methods and
Strategies of Web Personalization. Lecture Notes in Computer Science, Vol. 4321. Springer-
Verlag, Berlin Heidelberg New York (2007) this volume
126. Micarelli, A., Sciarrone, F.: A case-based system for adaptive hypermedia navigation. In:
Smith, I., Faltings, B. (eds.): Advances in Case-Based Reasoning. Lecture Notes in Artificial
Intelligence, Springer-Verlag, Berlin (1996) 266-279
127. Micarelli, A., Sciarrone, F.: Anatomy and empirical evaluation of an adaptive Web-based
information filtering system. User Modeling and User Adapted Interaction 14, 159-200 (2004)
128. Middleton, S.E., Shadbolt, N.R., De Roure, D.C.: Ontological User Profiling in
Recommender Systems. ACM Transactions on Information Systems 22, 1 (2004) 54-88
129. Millán, E., Agosta, J.M., Perez de la Cruz, J.-L.: Bayesian Student Modelling and the
Problem of Parameter Specification. British Journal of Educational Technology 32, 2 (2001)
171-181
130. Millán, E., Perez de la Cruz, J.-L.: A Bayesian Diagnostic Algorithm for Student
Modeling. User Modeling and User-Adapted Interaction 12 (2002) 281-330
131. Millán, E., Perez de la Cruz, J.-L., García, F.: Dynamic versus Static Student Models
Based on Bayesian Networks: An Empirical Study. In: Proc. of 7th International Conference
KES'2003. Lecture Notes in Computer Science, Vol. 2774. Springer-Verlag (2003) 1337-1344
132. Mislevy, R., Gitomer, D.H.: The Role o f Probability-Based Inference in an Intelligent
Tutoring System. User Modeling and User-Adapted Interaction 5, 3-4 (1996) 253-282
133. Mitchell, T., Chen, S.Y., Macredie, R.: Adapting Hypermedia to cognitive styles: Is it
necessary? In: Proc. of Workshop on Individual Differences in Adaptive Hypermedia at the 3rd
International Conference on Adaptive Hypermedia and Adaptive Web-based Systems. (2004)
http://www.dcs.bbk.ac.uk/~gmagoulas/AH2004_Workshop/Proceedings.htm
134. Mitrovic, A.: An Intellignet SQL Tutor on the Web. International Journal of Artificial
Intelligence in Education 13, 2-4 (2003) 173-197
135. Mobasher, B.: Data mining for Web personalization. In: Brusilovsky, P., Kobsa, A.,
Neidl, W. (eds.): The Adaptive Web: Methods and Strategies of Web Personalization. Lecture
Notes in Computer Science, Vol. 4321. Springer-Verlag, Berlin Heidelberg New York (2007)
this volume
136. Muldner, K., Conati, C.: Using Similarity to Infer Meta-Cognitive Behaviours During
Analogical Problem Solving. In: Proc. of 10th International Conference UM'05. Vol. 3538.
Springer-Verlag (2005) 134-143
137. Müller, C., Großmann-Hutter, B., Jameson, A., Rummer, R., Wittig, F.: Recognizing time
pressure and cognitive load on the basis of speech: An Experimental study. In: Bauer, M.,
Gmytrasiewicz, P.J., Vassileva, J. (eds.) Proc. of 8th International Conference on User Modeling,
UM 2001. Lecture Notes on Artificial Intelligence, Vol. 2109. Springer-Verlag (2001) 24-33
138. Murray, R.C., VanLehn, K., Mostow, J.: Looking ahead to select tutorial actions: A
decision-theoretic approach. International Journal of Artificial Intelligence in Education 14, 3-4
(2004) 235-279
139. Murray, W.: An Easily Implemented, Linear-time Algorithm for Bayesian Student
Modeling in Multi-level Trees. In: Lajoie, S., Vivet, M. (eds.) Proc. of 9th World Conference of
Artificial Intelligence and Education AIED'99. IOS Press (1999) 413-420
140. Neapolitan, R.: Probabilistic Reasoning in Expert Systems: Theory and Algorithms. John
Wiley & Sons, New York (1990)

141. Neumann, G., Zirvas, J.: SKILL - A scallable internet-based teaching and learning system.
In: Maurer, H., Olson, R.G. (eds.) Proc. of WebNet'98, World Conference of the WWW,
Internet, and Intranet. AACE (1998) 688-693
142. Not, E., Petrelli, D., Sarini, M., Stock, O., Strapparava, C., Zancanaro, M.:
Hypernavigation in the physical space: adapting presentation to the user and to the situational
context. New Review of Multimedia and Hypermedia 4 (1998) 33-45
143. Oberlander, J., O'Donell, M., Mellish, C., Knott, A.: Conversation in the museum:
experiments in dynamic hypermedia with the intelligent labeling explorer. The New Review of
Multimedia and Hypermedia 4 (1998) 11-32
144. Ohlsson, S .: Constraint-based student modeling. Journal of Artificial Intelligence in
Education 3, 4 (1992) 429-447
145. Okazaki, Y., Watanabe, K., Kondo, H.: An Implementation of the WWW Based ITS for
Guiding Differential Calculations. In: Brusilovsky, P., Nakabayashi, K., Ritter, S. (eds.) Proc. of
Workshop "Intelligent Educational Systems on the World Wide Web" at 8th World Conference
on Artificial Intelligence in Education. (1997) 18-25, also available at
http://www.contrib.andrew.cmu.edu/~plb/AIED97_workshop/Okazaki/Okazaki.html
146. Ong, E., Tay, A.-H., Ong, C.-K., Chan, S.-K.: Personalising Information Assets in
Collaborative Learning Environments. In: Looi, C.-K., McCalla, G., Bredeweg, B., Breuker, J.
(eds.) Proc. of 12th International Conference on Artificial Intelligence in Education, AIED'2005.
IOS Press (2005) 523-530
147. Ortony, A., Clore, G.L., Collins, A.: The Cognitive Structure of Emotions. Cambridge
University Press, (1988)
148. Papanikolaou, K.A., Grigoriadou, M., Kornilakis, H., Magoulas, G.D.: Personalising the
interaction in a Web-based Educational Hypermedia System: the case of INSPIRE. User
Modeling and User Adapted Interaction 13, 3 (2003) 213-267
149. Paris, C., Wan, S., Wilkinson, R., Wu, M.: Generating personal travel guides - and who
wants them? In: Bauer, M., Gmytrasiewicz, P.J., Vassileva, J. (eds.) Proc. of 8th International
Conference on User Modeling, UM 2001. Lecture Notes on Artificial Intelligence, Vol. 2109.
Springer-Verlag (2001) 251-253
150. Pask, G.: A fresh look at cognition and the individual. International Journal on the Man-
Machine Studies 4 (1972) 211-216
151. Paternò, F., Paganelli, L.: Intelligent analysis of user interactions with Web applications.
In: Gil, Y., Leake, D.B. (eds.) P roc. of 2002 International Conference on Intelligent User
Interfaces. ACM Press (2002) 111-118
152. Pazzani, M.J., Billsus, D.: Content-based recommendation systems. In: Brusilovsky, P.,
Kobsa, A., Neidl, W. (eds.): The Adaptive Web: Methods and Strategies of Web Personalization.
Lecture Notes in Computer Science, Vol. 4321. Springer-Verlag, Berlin Heidelberg New York
(2007) this volume
153. Pearl, J.: Probabilistic Reasoning in Expert Systems: Networks of Plausible Inference.
Morgan Kaufmann Publishers, Inc, San Francisco (1988)
154. Pentland, A.: Socially Aware Computation and Communication. Computer 38, 3 (2005)
33 - 40
155. Pérez, T., Gutiérrez, J., Lopistéguy, P.: An adaptive hypermedia system. In: Greer, J. (ed.)
Proc. of AI-ED'95, 7th World Conference on Artificial Intelligence in Education. AACE (1995)
351-358
156. Picard, R.W.: Affective Computing. MIT Press, Cambridge, MA (1997)

157. Pilar da Silva, D., Durm, R.V., Duval, E., Olivié, H.: Concepts and documents for
adaptive educational hypermedia: a model and a prototype. In: Brusilovsky, P., De Bra, P. (eds.)
Proc. of Second Adaptive Hypertext and Hypermedia Workshop at the Ninth ACM International
Hypertext Conference Hypertext'98. Eindhoven University of Technology (1998) 35-43, also
available as http://wwwis.win.tue.nl/ah98/Pilar/Pilar.html
158. Polson, M.C., Richardson, J.J. (eds.): Foundations of intelligent tutoring systems.
Lawrence Erlbaum Associates, Hillsdale (1988)
159. Prendinger, H., Mori, J., Ishizuka, M.: Recognizing, modeling, and responding to user
affective states. In: Ardissono, L., Brna, P., Mitrovic, A. (eds.) Proc. of 10th International User
Modeling Conference. Lecture Notes in Artificial Intelligence, Vol. 3538. Springer Verlag
(2005) 60-69
160. Prentzas, J., Hatzilygeroudis, I., Garofalakis, J.: A Web-based intelligent tutoring systems
using hybrid rules as its representation basis. In: Cerri, S.A., Gouardères, G., Paraguaçu, F. (eds.)
Proc. of 6th International Conference on Intelligent Tutoring Systems (ITS'2002). Lecture Notes
in Computer Science, Vol. 2363. Springer-Verlag (2002) 119-128
161. Read, T., Bárcena, E., Barros, B ., Verdejo, F .: I-PETER: Modelling Personalised
Diagnosis and Material Selection for an Online English Course. In: Proc. o f 8th Ibero-American
Conference IBERAMIA'2002. Lecture Notes in Computer Science, Vol. 2527. Springer-Verlag
(2002) 734-744
162. Reye, J.: Two-phase Updating of Student Models Based on Dynamic Belief Networks. In:
Proc. of 4th International Conference on Intelligent Tutoring Systems, ITS'98. Lecture Notes in
Computer Science, Vol. 1452. Springer-Verlag (1998) 6-15
163. Rich, E.: Building and Exploiting User Models. In: Proc. of Sixth International Joint
Conference on Artificial Intelligence. (1979) 720-722
164. Rich, E.A.: Stereotypes and user modeling. In: Kobsa, A., Wahlster, W. (eds.): User
models in dialog systems. Vol. 18. Springer-Verlag, Berlin (1989) 35-51
165. Riding, R., Rayner, S.: Cognitive Styles and Learning Strategies: Understanding Style
Differences in Learning and Behavior. David Fulton Publisher, London (1998)
166. Rist, T.: A perspective on intelligent information interfaces for mobile users. In: Smith,
M., Salvendy, G., Harris, D., Koubek, R.J. (eds.) Proc. of 9th International Conference on
Human-Computer Interaction, HCI International'2001. Vol. 1. Lawrence Erlbaum Associates
(2001) 154-158
167. Ritter, S.: PAT Online: A Model-tracing tutor on the World-wide Web. In: Brusilovsky,
P., Nakabayashi, K., Ritter, S. (eds.) Proc. of Wo rkshop "Intelligent Educational Systems on the
World Wide Web" at AI-ED'97, 8th World Conference on Artificial Intelligence in Education.
ISIR (1997) 11-17, also available at
http://www.contrib.andrew.cmu.edu/~plb/AIED97_workshop/Ritter/Ritter.html
168. Russell, S., Norvig, P.: Artificial Intelligence: A Modern Approach. Prentice Hall, (1995)
169. Sanrach, C., Grandbastien, M.: ECSAIWeb: A Web-based authoring system to create
adaptive learning systems. In: Brusilovsky, P., Stock, O., Strapparava, C. (eds.) Proc. of
Adaptive Hypermedia and Adaptive Web-based Systems, AH2000. Lecture Notes in Computer
Science, Vol. 1892. Springer-Verlag (2000) 214-226
170. Santos Jr., E., Nguyen, H., Zhao, Q., Wang, H.: User modeling for intent prediction in
information analysis. In: Proc. of 47th Annual Meeting for the Human Factors and Ergonomics
Society (HFES-03). (2003) 1034-1038
171. Sarini, M., Strapparava, C.: Building a User Model for a Mu seum Exploration and
Information-Providing Adaptive System. In: Brusilovsky, P., De Bra, P. (eds.) Proc. of Second

Adaptive Hypertext and Hypermedia Workshop at the Ninth ACM International Hypertext
Conference Hypertext'98. (1998) 63-68, also available at
http://wwwis.win.tue.nl/ah98/Sarini/Sarini.html
172. Schafer, J.B., Frankowski, D., Herlocker, J., Sen, S.: Collaborative filtering recommender
systems. In: Brusilovsky, P., Kobsa, A., Neidl, W. (eds.): The Adaptive Web: Methods and
Strategies of Web Personalization. Lecture Notes in Computer Science, Vol. 4321. Springer-
Verlag, Berlin Heidelberg New York (2007) this volume
173. Schmidt, A., Beigl, M., Gellersen, H.-W.: There is more to context than location.
Computers and Graphics 23, 6 (1999) 893-901
174. Schneider-Hufschmidt, M., Kühme, T., Malinowski, U. (eds.): Adaptive user interfaces:
Principles and practice. Human Factors in Information Technology, North-Holland, Amsterdam
(1993)
175. Sleeman, D.H.: UMFE: a user modeling front end system. International Journal on the
Man-Machine Studies 23 (1985) 71-88
176. Sosnovsky, S., Brusilovsky, P.: Layered Evaluation of Topic-Based Adaptation to Student
Knowledge. In: Proc. of Fourth Workshop on the Evaluation of Adaptive Systems at 10th
International User Modeling Conference, UM 2005. (2005) 47-56
177. Specht, M., Klemke, R.: ALE - Adaptive Learning Environment. In: Fowler, W.,
Hasebrook, J. (eds.) Proc. of WebNet'2001, World Conference of the WWW and Internet. AACE
(2001) 1155-1160
178. Specht, M., Kobsa, A.: Interaction of domain expertise and interface design in adaptive
educational hypermedia. In: Brusilovsky, P., De Bra, P. (eds.) Proc. of Second Workshop on
Adaptive Systems and User Modeling on the World Wide Web. (1999) 89-93
179. Specht, M., Oppermann, R.: ACE - Adaptive Courseware Environment. The New Review
of Hypermedia and Multimedia 4 (1998) 141-161
180. Stacey, K.P., Sonenberg, L., Nicholson, A., Boneh, T., S teinle, V.: A Teaching Model
Exploiting Cognitive Conflict Driven by a Bayesian Network. In: Brusilovsky, P., Corbett, A.,
Rosis, F.d. (eds.) Proc. of 9th International User Modeling Conference. Lecture Notes in
Computer Science, Vol. 2702. Springer-Verlag (2003) 352-362
181. Stash, N., Cristea, A., De Bra, P.: Authoring of learning styles in adaptive hypermedia:
Problems and solutions. In: Proc. of The 13th International World Wide Web Conference
(Alternate track papers and posters). ACM Press (2004) 114-123
182. Stathacopoulou, S., Magoulas, G., Grigoriadou, M., Samarakou, M.: Neuro-fuzzy
knowledge processing in intelligent learning environments for improved student diagnosis.
Information Sciences 170, 2-4 (2005) 273-307
183. Steinacker, A., Faatz, A., Seeberg, C., Rimac, I., Hörmann, S., Saddik, A.E., Steinmetz,
R.: MediBook: Combining semantic networks with metadata for learning resources to build a
Web based learning system. In: Proc. o f ED-MEDIA'2001 - World Conference on Educational
Multimedia, Hypermedia and Telecommunications. AACE (2001) 1790-1795
184. Steinacker, A., Seeberg, C., Rechenberger, K., Fischer, S., Steinmetz, R.: Dynamically
generated tables of contents as guided tours in adaptive hypermedia systems. In: Kommers, P.,
Richards, G. (eds.) Proc. of ED-MEDIA/ED-TELECOM'99 - 11th World Conference on
Educational Multimedia and Hypermedia and World Conference on Educational
Telecommunications. AACE (1999) 640-645
185. Stojanovic, L., Staab, S., Studer, R.: eLearning based on the Semantic Web. In: Fowler,
W., Hasebrook, J. (eds.) Proc. of WebNet'2001, World Conference of the WWW and Internet.
AACE (2001) 1774-1783

186. Suebnukarn, S., Haddawy, P.: Modeling Individual and Collaborative Problem Solving in
Medical Problem-Based Learning. In: Ardissono, L., Brna, P., Mitrovic, A. (eds.) Proc. of 10th
International User Modeling Conference, UM'2005. Lecture Notes in Artificial Intelligence, Vol.
3538. Springer-Verlag (2005) 377-386
187. Tanudjaja, F., Mui, L.: Persona: A contextualized and person alized Web search. In: Proc.
of 35th Hawaii International Conference on System Sciences. IEEE (2002) 1232-1240
188. Tarpin-Bernard, F., Habieb-Mammar, H.: Modeling elementary cognitive abilities for
adaptive hypermedia presentation. User Modeling and User Adapted Interaction 15, 5 (2005)
459-495
189. Trella, M., Carmona, C., Conejo, R.: MEDEA: an Open Service-Based Learning Platform
for Developing Intelligent Educational Systems for the Web. In: Proc. of Workshop on Adaptive
Systems for Web-based Education at 12th International Conference on Artificial Intelligence in
Education, AIED'2005. IOS Press (2005) 27-34
190. Triantafillou, E., Pomportis, A., Demetriadis, S .: The design and the formative evaluation
of an adaptive educational system based on cognitive styles. Computers and Education (2003)
87-103
191. Triantafillou, E., Pomportis, A., Demetriadis, S., Georgiadou, E.: The value of adaptivity
based on cognitive style: an empirical study. British Journal of Educational Technology 35, 1
(2004) 95–106
192. Tsiriga, V., Virvou, M.: Modelling the Student to Individualise Tutoring in a Web-Based
ICALL. International Journal of Continuing Engineering Education and Lifelong Learning 13, 3-
4 (2003) 350-365
193. Ueno, M.: Intelligent LMS with an agent that learns from log data. In: Richards, G. (ed.)
Proc. of World Conference on E-Learning, E-Learn 2005. AACE (2005) 2068-2074
194. VanLehn, K.: Student models. In: Polson, M.C., Richardson, J.J. (eds.): Foundations of
intelligent tutoring systems. Lawrence Erlbaum Associates, Hillsdale (1988) 55-78
195. VanLehn, K., Martin, J.: Evaluation of an assessment system based on Bayesian Student
Modeling. International Journal of Artificial Intelligence in Education 8, 2 (1998) 179-221
196. VanLehn, K., Niu, Z., Siler, S., Gertner, A.S.: Student Modeling from Conventional Test
Data: A Bayesian Approach Without Priors. In: Goettl, B., Redfield, C.L., Halff, H.M., Shute,
V.J. (eds.) Proc. of 4th International Conference on Intelligent Tutoring Systems, ITS'98. Lecture
Notes in Computer Science, Vol. 1452. Springer-Verlag (1998) 434-443
197. Vassileva, J.: An architecture and methodology for creating a domain-independent, plan-
based intelligent tutoring system. Educational and Training Technology International 27, 4
(1990) 386-397
198. Vassileva, J.: A task-centered approach for user modeling in a hypermedia office
documentation system. User Modeling and User-Adapted Interaction 6, 2-3 (1996) 185-224
199. Vassileva, J.: DCG + GTE: Dynamic Courseware Generation with Teaching Expertise.
Instructional Science 26, 3/4 (1998) 317-332
200. Vassileva, J., McCalla, G., Greer, J.: Multi-Agent Multi-User Modeling in I-Help. User
Modeling and User-Adapted Interaction 12 (2003) 179-210
201. Vomlel, J.: Bayesian Networks in Educational Testing. International Journal of
Uncertainty, Fuzziness and Knowledge-Based System 12 (2004) 83-100
202. W3C: Device Independence: Access to a Unified Web from Any Device in Any Context
by Anyone. World Wide Web Consortium (2006) http://www.w3.org/2001/di/
203. Weber, G., Brusilovsky, P.: ELM-ART: An adaptive versatile system for Web-based
instruction. International Journal of Artificial Intelligence in Education 12, 4 (2001) 351-384

204. Weber, G., Kuhl, H.-C., Weibelzahl, S.: Developing adaptive internet based courses with
the authoring system NetCoach. In: Bra, P.D., Brusilovsky, P., Kobsa, A. (eds.) Proc. of Third
workshop on Adaptive Hypertext and Hypermedia. (2001) 35-48, also available at
http://wwwis.win.tue.nl/ah2001/papers/GWeber-UM01.pdf
205. Witkin, H.A., Moore, C.A., Goodenough, D.R., Cox, P.W.: Field-dependent and field-
independent cognitive styles and their educational implications. Review of Educational Research
47, 1 (1977) 1-64
206. Yin, X., Lee, W.S., Tan, Z.: Personalization of Web Content for Wireless Mobile Device.
In: Proc. of Wireless Communications and Networking Conference. IEEE (2004) 2569-2574
207. Zadeh, L.: Fuzzy sets. Information and Control 8 (1965) 338-353
208. Zapata-Rivera, D., Greer, J.: SModel Server: Student Modelling in Distributed Multi-
Agent Tutoring Systems. In: Moore, J.D. (ed.) Proc. of 9th World Conference of Artificial
Intelligence and Education, AIED'99. IOS Press (2001) 446-455
209. Zapata-Rivera, D., Greer, J.: Inspectable Bayesian student modelling servers in multi-
agent tutoring systems. International Journal of Human-Computer Studies 61 (2004) 535-563
210. Zimmermann, A., Specht, M., Lorenz, A.: Personalization and Context Management. User
Modeling and User-Adapted Interaction 15, 3-4 (2005) 275-302