Content uploaded by Roberto Saia
Author content
All content in this area was uploaded by Roberto Saia on Oct 11, 2017
Content may be subject to copyright.
Binary Sieves: Toward a Semantic Approach to User
Segmentation for Behavioral Targeting
Roberto Saia, Ludovico Boratto, Salvatore Carta, and Gianni Fenu1
Dipartimento di Matematica e Informatica, Universit`a di Cagliari
Via Ospedale 72, 09124 Cagliari, Italy
Abstract
Behavioral targeting is the process of addressing ads to a specific set of users.
The set of target users is detected from a segmentation of the user set, based
on their interactions with the website (pages visited, items purchased, etc.).
Recently, in order to improve the segmentation process, the semantics behind
the user behavior has been exploited, by analyzing the queries issued by the
users. However, nearly half of the times users need to reformulate their queries
in order to satisfy their information need. In this paper, we tackle the problem
of semantic behavioral targeting considering reliable user preferences, by per-
forming a semantic analysis on the descriptions of the items positively rated by
the users. We also consider widely-known problems, such as the interpretability
of a segment, and the fact that user preferences are usually stable over time,
which could lead to a trivial segmentation. In order to overcome these issues,
our approach allows an advertiser to automatically extract a user segment by
specifying the interests that she/he wants to target, by means of a novel boolean
algebra; the segments are composed of users whose evaluated items are seman-
tically related to these interests. This leads to interpretable and non-trivial
segments, built by using reliable information. Experimental results confirm the
effectiveness of our approach at producing users segments.
Keywords: User Segmentation, Semantic Analysis, Behavioral Targeting
Email address: {roberto.saia,ludovico.boratto,salvatore,fenu}@unica.it (Roberto
Saia, Ludovico Boratto, Salvatore Carta, and Gianni Fenu)
Preprint submitted to Future Generation Computer Systems May 26, 2016
1. Introduction
Behavioral targeting addresses ads to a set of users who share common prop-
erties. In order to choose the set of target users that will be advertised with a
specific ad, a segmentation that partitions the users and identifies groups that
are meaningful and different enough is first performed. In the literature it has
been highlighted that classic approaches to segmentation (like k-means) cannot
take into account the semantics of the user behavior [1]. Tu and Lu [2] proposed
a user segmentation approach based on a semantic analysis of the queries issued
by the users, while Gong et al. [1] proposed a LDA-based semantic segmentation
that groups users with similar query and click behaviors.
When dealing with a semantic behavioral targeting approach, several prob-
lems remain open.
Reliability of a semantic query analysis. In the literature it has been highlighted
that half of the time users need to reformulate their queries, in order to satisfy
their information need [3, 4, 5]. Therefore, the semantic analysis of a query is not
a reliable source of information, since it does not contain any information about
whether or not a query led to what the user was really looking for. Moreover,
performing a semantic analysis on the items evaluated by the users in order to
filter them can increase the accuracy of a system [6, 7, 8]. Therefore, a possible
way to overcome this issue would be to perform a semantic analysis on the
description of the items a user positively evaluated through an explicitly given
rating. However, another issue arises in cascade.
Preference stability. To complicate the previous scenario, there are domains
like movies in which the preferences tend to be stable over time [9] (i.e., users
tend to watch movies of the same genres or by the same director/actor). This
is useful to maintain high-quality knowledge sources, but considering only the
items a user evaluated leads to trivial sets of users that represent the target
(this problem is known as overspecialization [10]).
2
Interpretability of the segments. The last open problem that has to be faced in
this research area is the interpretability of a segment. Indeed, a recent survey
on user segmentation (mostly focused on the library domain) [11], highlighted
that, in order to create a proper segmentation of the users, it is important to
understand them. On the one hand, easily interpretable approaches generate
trivial segments, and even a partitioning with the k-means clustering algorithm
has proven to be more effective than this method [12], while on the other hand,
when a larger set of features is combined, the problem of properly understanding
and interpreting results arises [13, 14]. This is mostly due to the lack of guid-
ance on how to interpret the results of a segmentation [15]. The fact that easily
understandable approaches generate ineffective segments, and that more com-
plex ones are accurate but not easy to use in practice, generates an important
gap in this research area.
Our contributions. In this paper, we have moved the item analysis process from
the canonical deterministic space model (i.e., that based on strict mathematical
criteria) to a more flexible semantic space model that allows us to extend the
analysis capability, which in the literature has been highlighted as a challenging
topic [16, 17]. In particular, we tackle the problem of defining a semantic user
segmentation approach, such that the sources of information used to build it are
reliable, the generated segmentation is not trivial and it is easily interpretable.
The proposed approach is based on a semantic analysis of the description of
the items positively evaluated by the users. The choice to start from items with
a positive score was made since it is necessary to start from a knowledge-base
that accurately describes what the users like, so that our approach can employ
the semantics to detect latent information and avoid preference stability.
The approach first defines a binary filter (called semantic binary sieve) for
each class of items that, by analyzing the description of the items classified
with the class, defines which words characterize it. In order to detect more
complex targets, we are going to define an algorithm that takes as input a set of
classes that characterize the ads that have to be proposed to the users and a set
3
of boolean operators. The algorithm combines the classes with the operators
by means of a boolean algebra, and creates the binary filters that characterize
the combined classes. Then we consider the words (that as we will explain
later, are actually particular semantic entities named synsets) that describe the
items evaluated by a user, and use the previously created filters to evaluate a
relevance score that indicates how relevant is each class of items for the user.
The relevance scores of each user are filtered by the segmentation algorithm, in
order to return all the users characterized by a specified class or set of classes.
By selecting segments of users who are semantically related to the classes
specified by the advertisers, we avoid considering only the users who evaluated
items of that class; this allows our approach to overcome the open problems pre-
viously mentioned, related to preference stability and to the triviality of a seg-
mentation generated by considering the evaluated items. Moreover, by defining
the semantic binary sieves that characterize each class and the relevance scores
that characterize each user, we avoid the interpretability issues that usually af-
fect the user segmentation; indeed, each class of items is described by thousands
of features (i.e., the words that characterize it), but this complexity is hidden to
the advertiser, which is only required to specify the users she/he wants to target
(e.g., those whose models are characterized by comedy AND romantic movies).
Considering that the evaluation of the users for the items offered in a context
of e-commerce are usually thousands or millions, the proposed approach repre-
sents an efficient strategy to model in a compact way the information related
to these big amounts of data.
The scientific contributions of our proposal are now recapped:
•we introduce a novel data structure, called semantic binary sieve, to se-
mantically characterize each class of items;
•we present a semantic user segmentation approachbased on reliable sources
of information; with respect to the state-of-the-art approaches that are
based on the semantic analysis of the queries issued by the users, we
perform a semantic analysis on the description of the items positively
4
evaluated by the users;
•we solve the overspecialization issues caused by preference stability by
building a model for each user that considers her/him as interested in a
class of items, if the items she/he evaluated are semantically related with
the words that characterize that class;
•we present a boolean algebra that allows us to specify, in a simple but
punctual way, the interests that the segment should cover; this algebra,
along with the built models, avoids the interpretability issues that usually
characterize the segmentations built with several features;
•we perform five sets of experiments on a real-world dataset, with the aim to
validate our proposal by analyzing the different ways in which the classes
can be combined through the boolean operations. The generated segments
will be evaluated by comparing them with the topic-based segmentation
(as several state-of-the-art approaches do), based on the real choices of
the users.
The rest of the paper is organized as follows: we first present the works in the
literature related with our approach (Section 2), then we provide a background
on the concepts handled by our proposal and the formal definition of the tackled
problem (Section 3), we continue with the implementation details (Section 4)
and the description of the performed experiments (Section 5), ending with some
concluding remarks (Section 6).
2. Related Work
In this section we are going to explore the main works in the literature related
to the open problems highlighted in the Introduction.
Behavioral targeting. A high variety of behavioral targeting approaches has
been designed by the industry and developed as working products. Google’s
5
AdWords1performs different types of targeting to present ads to users; the clos-
est to our proposal is the “Topic targeting”, in which the system groups and
reaches the users interested in a specific topic. DoubleClick2is another system
employed by Google that exploits features such as browser information and the
monitoring of the browsing sessions. In order to reach segments that contain
similar users, Facebook offers Core Audiences3, a tool that allows advertisers to
target users with similar location, demographic, interests, or behaviors; in par-
ticular, the interest-based segmentation allows advertisers to choose a topic and
target a segment of users interested by it. Among its user targeting strategies,
Amazon offers the so-called Interest-based ads policy4, a service that detects
and targets segments of users with similar interests, based on what the users
purchased, visited, and by monitoring different forms of interaction with the
website (e.g., the Amazon Browser Bar). SpecificMedia5uses anonymous web
surfing data in order to predict a user’s purchase prediction score. Yahoo! Be-
havioral Targeting6creates a model with the online interactions of the users,
such as searches, page-views, and ad interactions to predict the set of users
to target. Other commercial systems, such as Almond Net7,Burst8,Phorm9,
and Revenue Science10 include behavioral targeting features. Research studies,
such as the one presented by Yan et al. [18], show that an accurate monitoring
of the click-through log of advertisements collected from a commercial search
engine can help online advertising. Beales [19] collected data from online adver-
tising networks and showed that a behavioral targeting performed by exploiting
prices and conversion rates (i.e., the likelihood of a click to lead to a sale) is
1https://support.google.com/adwords/answer/1704368?hl=en
2https://www.google.com/doubleclick/
3https://www.facebook.com/business/news/Core-Audiences
4http://www.amazon.com/b?node=5160028011
5http://specificmedia.com/
6http://advertising.stltoday.com/content/behavioral FAQ.pdf
7http://www.almondnet.com/
8http://www.burstmedia.com/
9http://www.phorm.com/
10http://www.revenuescience.com/
6
twice more effective than traditional advertising. Chen et al. [20] presented a
scalable approach to behavioral targeting, based on a linear Poisson regression
model that uses granular events (such as individual ad clicks and search queries)
as features. Approaches to exploit the semantics [6, 7] or the capabilities of a
recommender system [21, 22, 23] to improve the effectiveness of the advertising
have been proposed, but none of them generates segments of target users.
Segment interpretability and semantic user segmentation. Choosing the right
criteria to segment users is a widely studied problem in the market segmenta-
tion literature, and two main classes of approaches exist. On the one hand, the a
priori [24] or commonsense [25] approach is based on a simple property, like the
age, which is used to segment the users. Even though the generated segments
are very easy to understand and they can be generated at a very low cost, the
segmentation process is trivial and even a partitioning with the k-means clus-
tering algorithm has proven to be more effective than this method [12]. On
the other hand, post hoc [26] approaches (also known as a posteriori [24] or
data-driven [25]) combine a set of features (which are known as segmentation
base [27]) in order to create the segmentation. Even though these approaches
are more accurate when partitioning the users, the problem of properly under-
standing and interpreting results arises [13, 14]. This is mostly due to the lack
of guidance on how to interpret the results of a segmentation [15]. Regarding
the literature on behavioral user segmentation, Bian et al. [28] presented an
approach to leverage historical user activity on real-world Web portal services
to build a behavior-driven user segmentation. Yao et al. [29] adopted SOM-
Ward clustering (i.e., Self Organizing Maps, combined with Ward clustering),
to segment a set of customers based on their demographic and behavioral char-
acteristics. Zhou et al. [30] performed a user segmentation based on a mixture
of factor analyzers (MFA) that consider the navigational behavior of the user in
a browsing session. Regarding the semantic approaches to user segmentation,
Tu and Lu [2] and Gong et al. [1] both proposed approaches based on a semantic
analysis of the queries issued by the user through Latent Dirichlet Allocation-
7
based models, in which users with similar query and click behaviors are grouped
together. Similarly, Wu et al. [31] performed a semantic user segmentation by
adopting a Probabilistic Latent Semantic Approach on the user queries. As
this analysis showed, none of the behavioral targeting approaches exploits the
interactions of the users with a website in the form of a positive rating given to
an item.
Preference stability. As mentioned in the Introduction, Burke and Ramezani
highlighted that some domains are characterized by a stability of the preferences
over time [9]. Preference stability leads also to the fact that when users get in
touch with diverse items, diversity is not valued [32]. On the one side, users tend
to access to agreeable information (a phenomenon known as filter bubble [33])
and this leads to the overspecialization problem [10], while on the other side
they do not want to face diversity. Another well-known problem is the so called
selective exposure, i.e., the tendency of users to make their choices (goods or
services) based only on their usual preferences, which excludes the possibility for
the users to find new items that may be of interest to them [34]. The literature
presents several approaches that try to reduce this problem, e.g., NewsCube [35]
operates by offering to the users several points of view, in order to stimulate
them to make different and unusual choices.
3. Preliminaries
Background. For many years the item descriptions were analyzed through a
word vector space model, where all the words of each item description are pro-
cessed by TF-IDF [36] and stored in a weighted vector of words.
Due to the fact that this approach based on a simple bag of words is not
able to perform a semantic disambiguation of the words in an item description,
because it does not adopt any semantic data model [37], and motivated by
the fact that the exploitation of a taxonomy for categorization purposes is an
approach recognized in the literature [38], we decided to use the functionalities
offered by the WordNet environment. Wordnet is a large lexical database of
8
English, where nouns,verbs,adjectives, and adverbs are grouped into sets of
cognitive synonyms (synsets), each expressing a distinct concept. Synsets are
interlinked by means of conceptual-semantic and lexical relations. Wordnet
currently contains about 155287 words, organized into 117659 synsets, for a total
of 206941 word-sense pairs [39]. In a short, the main relation among words in
WordNet is the synonymy and the synsets are unordered sets of grouped words
that denote the same concept and are interchangeable in many contexts. Each
synset is linked to other synsets through a small number of conceptual relations.
Word forms with several distinct meanings are represented in as many distinct
synsets, in this way each form-meaning pair in WordNet will be unique (e.g., the
fly noun and the fly verb belong to two distinct synsets). Most of the WordNet
relations connect words that belong to the same part-of-speech (POS). There
are four POS: nouns,verbs,adjectives, and adverbs. Both nouns and verbs are
organized into precise hierarchies, defined by a hypernym or is-a relationship.
For example, the first sense of the word radio would have the following hypernym
hierarchy, where the words at the same level are synonyms of each other: as
shown in the following, some sense of radio is synonymous with some other
senses of radiocommunication or wireless, and so on.
1. POS=noun
(a) radio, radiocommunication, wireless (medium for communication)
(b) radio receiver, receiving set, radio set, radio, tuner, wireless (an elec-
tronic receiver that detects and demodulates and amplifies transmitted
signals)
(c) radio, wireless (a communication system based on broadcasting elec-
tromagnetic waves)
2. POS=verb
(a) radio (transmit messages via radio waves)
We use the synsets to perform both the definition of binary filters and the
evaluation of the relevance scores of the classes in a user profile.
9
Problem Definition. Here, we define the problem handled by our proposal. A
set of definitions will first allow us to introduce the notation used in the problem
statement.
Definition 3.1 (User preferences). We are given a set of users U={u1,...,uN},
a set of items I={i1,...,iM}, and a set Vof values used to express the
user preferences (e.g., V= [1,5] or V={like, dislike}). The set of all pos-
sible preferences expressed by the users is a ternary relation P⊆U×I×V.
We denote as P+⊆Pthe subset of preferences with a positive value (i.e.,
P+={(u, i, v)∈P|v≥v∨v=like}), where vindicates the mean value (in
the previous example, in which V= [1,5],v= 3).
Definition 3.2 (User items and classes). Given the set of positive prefer-
ences P+, we denote as I+={i∈I| ∃(u, i, v)∈P+}the set of items for which
there is a positive preference, and as Iu={i∈I| ∃(u, i, v)∈P+∧u∈U}the
set of items a user ulikes. Let C={c1,...,cK}be a set of primitive classes
used to classify the items; we denote as Ci⊆Cthe set of classes used to classify
an item i(e.g., Cimight be the set of genres that a movie iwas classified with),
and with Cu={c∈C| ∃(u, i, v)∈P+∧i∈Ci}the classes associated to the
items that a user likes.
Definition 3.3 (Semantic item description). Let BoW ={t1,...,tW}be
the bag of words used to describe the items in I; we denote as dithe binary
vector used to describe each item i∈I(each vector is such that |di|=|BoW |).
We define as S={s1,...,sW}the set of synsets associated to BoW (that is,
for each word used to describe an item, we consider its associated synset), and
as sdithe semantic description of i. The set of semantic descriptions is denoted
as D={sd1, . . . , sdM}(note that we have a semantic description for each item,
so |D|=|I|). The approach used to extract sdifrom diis described in detail
in Section 4.1.
Definition 3.4 (Semantic Binary Sieve). Let Dc⊆Cbe the subset of se-
mantic descriptions of the items classified with a class c∈C(i.e., Dc={sdi|
10
c∈Ci}). We define as Semantic Binary Sieve (SBS), a binary vector bcthat
contains which synsets characterize that class. The algorithm to build a seman-
tic binary sieve is given in Section 4.3.
Definition 3.5 (Boolean class). Given the set of classes Cand a set of boolean
operators τ={∧,∨,¬}, a boolean class is a subset of Qclasses CQ⊆Ccom-
bined through a subset of boolean operators τQ⊆τ. A boolean class is repre-
sented as a semantic binary sieve that defines which synsets characterize the
combined classes. The algorithm to build the semantic binary sieve of a boolean
class is also given in Section 4.3.
Definition 3.6 (User segment). Given a set of users Uand a (boolean) class
cq, a user segment is a subset of users to target T⊆Uwhose positively evaluated
items Iuare semantically related to the items that belong to cq.
Problem 1. Given a set of positive preferences P+that characterizes the items
each user likes, a set of classes Cused to classify the items (possibly combined
with a set of boolean operators τ), and a set of semantic descriptions D, our
first goal is to assign a relevance score ru(c)for each user uand each class
c, based on the semantic descriptions D. The objective of our approach is to
define a function f:CK×τ→Uthat, given a (boolean) class, returns a set of
users (user segment) T⊆U, such that ∀u∈T , ru(c)≥ϕ(where ϕindicates a
threshold that defines when a score is relevant enough for the user to be included
in the target).
4. Applied Strategy
In this section we present our strategy, which performs a semantic analysis
of the descriptions of the items the users like, in order to model both the users
and the classes, and perform the semantic segmentation on the user set. Our
approach performs five steps:
1. Text preprocessing: processing of the textual information related to all
the items, in order to retrieve the synsets;
11
2. User Modeling: creation of a model that contains which synsets are
present in the items a user likes;
3. Semantic Binary Sieve definition: creation of the Semantic Binary
Sieves (SBS), i.e., a series of binary filters able to estimate which synsets
are relevant for a class; a class can either be a class with which an item
was classified, or a boolean class that combines primitive classes through
boolean operators (as primitive classes we mean the native classification
of the items present in the used dataset);
4. Relevance score definition: generation of a relevance score that allows
us to weight the user preferences in terms of classes;
5. Segment Definition: selection of the users characterized by a class or a
boolean class.
In the following, we describe in detail how each step works.
4.1. Text Preprocessing
Before extracting the WordNet synsets from the text that describes each
item, we need to follow several preprocessing tasks. The first is to detect the
correct Part-Of-Speech (POS) for each word in the text; in order to perform
this task, we have used the Stanford Log-linear Part-Of-Speech Tagger [40].
Then, we remove punctuation marks and stop-words, which represent noise in
the semantic analysis (in this work we have used a list of 429 stop-words made
available with the Onix Text Retrieval Toolkit11). After we have determined the
lemma of each word using the Java API implementation for WordNet Searching
JAWS12, we perform the so-called word sense disambiguation, a process where
the correct sense of each word is determined, which permits us to individuate the
appropriate synset. The best sense of each word in a sentence was found using
the Java implementation of the adapted Lesk algorithm provided by the Den-
mark Technical University similarity application [41]. All the collected synsets
11http://www.lextek.com/manuals/onix/stopwords.html
12http://lyle.smu.edu/ tspell/jaws/index.html
12
form the set S={s1,...,sW}defined in Section 3. The output of this step
is the semantic disambiguation of the textual description of each item i∈I,
which is stored in a binary vector dsi; each element of the vector dsi[w] is 1 if
the corresponding synset is a part of the item description, or it is 0 otherwise.
4.2. User Modeling
For each user u∈U, this step considers the set of items Iushe/he likes,
and builds a user model muthat describes which synsets characterize the user
profile (i.e., which synsets appear in the semantic description of these items).
Each model muis a binary vector that contains an element for each synset
sw∈S. In order to build the vector, we consider the semantic description dsi
of each item i∈Iufor which the user expressed a positive preference. In order
to build mu, this step performs the following operation on each element w:
mu[w] =
1, if dsi[w] = 1
mu[w], otherwise
(1)
This means that if the semantic description of an item icontains the synset
sw, the synset becomes relevant for the user, and we set to 1 the bit at position
win the user model mu; otherwise, its value remains unaltered. By performing
this operation for all the items i∈Iu, we model which synsets are relevant for
the user. The output of this step is a set M={m1,...,mN}of user models
(note that we have a model for each user, so |M|=|U|).
4.3. Semantic Binary Sieve Definition
Given a set of classes C, in this step we define a binary vector, called Se-
mantic Binary Sieve (SBS), which describes the synsets that characterize each
class. Moreover, we are going to present an approach to build the boolean classes
previously defined, i.e., a semantic binary sieve that describes multiple classes
combined through a set of boolean operators τ={∧,∨,¬}.
Therefore, four types of semantic binary sieves can be defined:
13
1. Primitive class-based SBS definition. Given a primitive class of items
ck, this operation creates a binary vector that contains the synsets that
characterize the description of the items classified with ck.
2. Interclass-based SBS definition. Given two classes ckand cq, we com-
bine the SBSs of the two classes with an AN D operator, in order to build
a new semantic binary sieve that contains the synsets that characterize
both the classes.
3. Superclass-based SBS definition. Given two classes ckand cq, we
combine the SBSs of the two classes with an OR operator, in order to
build a new semantic binary sieve that merges their synsets.
4. Subclass-based SBS definition. Given two classes ckand cq, we use the
SBS of cqas a binary negation mask on the SBS of ck, in order to build a
new semantic binary sieve that contains the synsets that characterize the
first class but do not characterize the second.
4.3.1. Primitive class-based SBS Definition
For each class ck∈C, we create a binary vector that stores which synsets
are relevant for that class. These vectors, called Semantic Binary Sieves, will
be stored in a set B={b1,...,bK}(note that |B|=|C|, since we have a
vector for each class). Each vector bk∈Bcontains an element for each synset
sw∈S(i.e., |bk|=|S|). In order to build the vector, we consider the semantic
description dsiof each item i∈I+for which there is a positive preference, and
each class ckwith whom iwas classified. The binary vector bkstores which
synsets are relevant for a class ck, by performing the following operation on
each element bk[w] of the vector:
bk[w] =
1, if dsi[w] = 1 ∧i∈ck,∀i∈I+
bk[w], otherwise
(2)
In other words, if the semantic description of an item icontains the synset sw,
the synset becomes relevant for each class ckthat classifies i, and the semantic
binary sieve bkassociated to ckhas the bit at position wset to 1; otherwise, its
14
value remains unaltered. By performing this operation for all the items i∈I+
that are classified with ck, we know which synsets are relevant for the class.
After we processed all the classes c∈Cwe obtain a description of the primitive
classes that allow us to build the filters for the boolean class.
4.3.2. Interclass-based SBS Definition
Starting from the set B={b1,...,bK}, we can arbitrarily manage the ele-
ments bk∈Bto generate boolean classes, i.e., a combination of primitive classes
by means of a boolean operator. The first type of boolean class we are going
to define, named interclass is formed by the combination of the binary sieves of
the two classes bkand bqthrough an AN D operator. Considering each element
wof the two vectors, which indicates if a synset wis relevant or not for a class,
the semantics of the operator is the following:
bk[w]∧bq[w] =
1, if bk[w] = 1 and bq[w] = 1
0, otherwise
(3)
This boolean class indicates which synsets characterize all the classes of
items involved. We can obtain this result by recurring to the axiomatic set
theory (i.e., the elementary set theory based on the Venn diagrams); indeed,
we can consider each class of items as a set, and create a new interclass that
characterizes the common elements of two or more SBSs, using an intersection
operation ∩;
The example in Figure 1 is a simple demonstration of what said based on the
axiomatic set theory. It describes the effect of a boolean AND operation applied
to the classes C1,C2, and C3: in this case the result of operation C1∩C2∩C3
represents a new interclass that we can use to refer to a precise segment of users,
in a more atomic way than with the use of the primitive classes.
To provide a more specific presentation of what is the result of an interclass-
based SBS, we are going to provide an example (presented in Table 1), in which
the two classes with most items in the dataset employed in our experiments
(i.e., the classes 1 and 5) are combined with an AND operator. In the example,
15
C1
C1
C2
C3
C1∩C2C2∩C3
C1∩C3
C1∩C2∩C3
Figure 1: Inter-class definition
Class Num. of 1 occurrences Num. of 0 occurrences % of 1 occurrences % of 0 occurrences
1 14175 6947 67.11 32.89
5 14825 6297 70.19 29.81
1 AND 5 11338 9784 53.68 46.32
Table 1: Example of interclass-based SBS considering the two classes with most items
the vector has a fixed length and contains 21122 elements, which represent the
synsets extracted from the dataset. The results show that when two classes
are combined in order to extract the synsets that characterize both, around
15% of synsets that characterize just one class are discarded by the resulting
interclass-based SBS. In other words, this SBS has more non-relevant synsets
with respect to the original classes (this is represented by the percentage of
zero occurrences), and provides knowledge of which synsets are able to describe
both classes of items, allowing a more specific and narrow user segmentation
that captures which users are interested in both classes.
4.3.3. Superclass-based SBS Definition
By combining the binary sieves of the two classes bkand bqthrough an
OR operator, we can generate a new type of boolean class, named superclass.
Considering each element wof the two vectors, which indicates if a synset wis
16
C1
C2
C3
C1
C2
C3
Figure 2: Superclass definition
relevant or not for a class, the semantics of the operator is the following:
bk[w]∨bq[w] =
1, if bk[w] = 1 or bq[w] = 1
0, otherwise
(4)
This boolean class would allow an advertiser to broaden a target, capturing
in a semantic binary sieve the synsets that are characterizing for two or more
classes. By using the axiomatic set theory, we can consider each class of items
as a set, and create a new superclass that characterizes more primitive classes
through a union operation ∪of two or more SBSs.
The example in Figure 2 shows a demonstration of what said, based on the
axiomatic set theory. It describes the effect of a boolean OR operation applied
to the classes C1,C2, and C3(represented by the grey area).
To provide a more specific presentation of what is the result of a superclass-
based SBS, Table 2 shows an example in which classes 1 and 5 are combined
with an OR operator. The results show that when two classes are combined in
order to extract the synsets that characterize both, around 15% of synsets that
characterize just one class are added to the resulting superclass-based SBS. In
other words, this SBS has less non-relevant synsets with respect to the original
17
Class Num. of 1 occurrences Num. of 0 occurrences % of 1 occurrences % of 0 occurrences
1 14175 6947 67.11 32.89
5 14825 6297 70.19 29.81
1 OR 5 17662 3460 83.62 16.38
Table 2: Example of superclass-based SBS considering the two classes with most items
classes (this is represented by the percentage of zero occurrences), and provides
knowledge of which synsets are able to describe at least one of the classes of
items, allowing a more broad user segmentation that captures which users are
interested in at least one of the classes.
4.3.4. Subclass-based SBS Definition
Another important entity that we can obtain through the managing of the
elements b∈Bis the subset of a primitive class. It means that we can extract
from a semantic binary sieve a subset of elements that express an atomic char-
acteristic of the source set. For instance, if we consider a dataset where the
items are movies, from a genre of classification we can extract several semantic
binary sieves that characterize some sub-genres of movies.
More formally, a subclass is a partition of a primitive or boolean class, e.g.,
for the primitive class Comedy we can define an arbitrary number of subclasses,
applying some operation of the axiomatic set theory. In the example in Figure 3,
we define a subclass Comedy \Romance, in which all the synsets that charac-
terize the Romance class are removed from the Comedy class. Therefore, only
the comedy movies that do contain romance elements are represented through
this boolean class.
Given two semantic binary sieves bkand bq, we can use bqas a binary
negation mask. For each element wof the vector, this operation modifies the
binary value of the destination bits, as shown in Equation (5).
bk[w] =
bk[w], if bq[w] = 0
0, otherwise
(5)
To provide a more specific presentation of what is the result of a subclass-
based SBS, we are going to provide an example (presented in Table 3), in which
18
Comedy
Comedy \Romance
Figure 3: Sub-class definition
Class Num. of 1 occurrences Num. of 0 occurrences % of 1 occurrences % of 0 occurrences
5 14825 6297 70.19 29.81
14 8853 6947 67.11 32.89
5 NOT 14 11338 12269 41.91 58.09
Table 3: Example of interclass-based SBS considering the two classes with most items
we combine with a NOT operator the two classes of the dataset that have been
most used to co-classify the items (i.e., the classes 5 and 14). The results show
that, when two classes are combined, around 30% of synsets that characterize
the first class are discarded by the resulting subclass-based SBS. In other words,
this SBS has more non-relevant synsets with respect to the first class from which
we removed the synsets that are relevant for the second, and provides knowledge
of which synsets describe the first class of items but not the second, allowing
a more specific and narrow user segmentation that captures which users are
interested in items of the first class that do not contain in their description
synsets of the second class.
4.3.5. Additional Considerations on the Boolean Classes
Given the elementary boolean operations we presented to create a boolean
class, given two classes and an operator, we can also create a new boolean class
using the results of the previous operations, by combining them with further
19
operations of the same type, e.g., (b1∨b2)∧(b2¬b3).
It should be also noted that only the NOT operation, together with one
of the other two operations (AND and OR) is enough to express all possible
combination of classes, as shown in Equation (6).
x∧y=¬(¬x∨ ¬y)
x∨y=¬(¬x∧ ¬y)
(6)
4.4. Relevance Score Definition
This step compares the output of the two previous steps (i.e., the set Bof
binary vectors related to the Semantic Binary Sieves, and the set Mof binary
vectors related to the user models), in order to infer which classes are relevant
for a user. The main idea is to consider which synsets are relevant for a user u
(this information is stored in the user model mu) and evaluate which classes are
characterized by the synsets in mu(this information is contained in each vector
bk, which contains the synsets that are relevant for the class ck). The objective
is to build a relevance score ru[k] that indicates the relevance of the class ck
for the user u. The key concept behind this step is that we do not consider the
items a user evaluated anymore. Each vector in Bis used as a filter (this is why
the vectors are called semantic binary sieves), and this allows us to estimate
the relevance of each class for that user. Therefore, the relevance score of a
class for a user can be used to generate non trivial segments, since a user might
be associated to classes of items she/he never expressed a preference for, but
characterized by synsets that also characterize the user model. By considering
each semantic binary sieve bk∈Bassociated to the class ckand the user model
mu, we define a matching criteria Θ between each synset mu[w] in the user
model, and the corresponding synset bk[w] in the semantic binary sieve, by
adding 1 to the relevance score of that class for the user (element ru[k]), if the
synset is set to 1 both in the semantic binary sieve and in the user model, and
leaving the current value as it is otherwise. The semantics of the operator is
20
shown in Equation (7).
bk[w]Θmu[w] =
ru[k] + +, if mu[w] = 1 and bk[w] = 1
ru[k], otherwise
(7)
The relevance scores built by this step will be used by our target definition
algorithm, in order to infer which users are characterized by a specific class or
set of classes.
4.5. Segment Definition
This step defines the set of users that are part of the target. Given a boolean
class of items c, we build a function f:CK×τ→U, that evaluates the relevance
score ru(c) of each user u∈Ufor that class, in order to understand if the class
is relevant enough for a user to be included in the target. More specifically, the
function operates as follows:
f(c) = {u∈U|ru(c)≥ϕ}(8)
where ϕis a threshold that defines the minimum value that the score has to
take in order to consider the user as relevant for the target.
5. Experiments
This section describes the experiments performed to validate our proposal.
In Section 5.1 we describe the experimental setup and strategy, in Section 5.2
the dataset employed for the evaluation is presented, Section 5.3 illustrates the
metrics, and Section 5.4 contains the results.
5.1. Experimental Setup and Strategy
The experiments have been performed using the Java language with the
support of Java API implementation for WordNet Searching (JAWS), and the
real-world dataset Yahoo! Webscope Movie dataset (R4)13. The experimental
13http://webscope.sandbox.yahoo.com
21
framework was developed by using a machine with an Intel i7-4510U, quad core
(2 GHz ×4) and a Linux 64-bit Operating System (Debian Jessie) with 4 GBytes
of RAM. To validate our proposal, we performed five sets of experiments:
1. Data overview. This experiment studies the distribution of the classes,
by considering for how many users each class is the most relevant (i.e.,
the one for which a user has given most positive ratings), in order to
evaluate how trivial it is to perform a segmentation based on the classes;
we also analyze the number of genres with which each item is classified, in
order to evaluate the capability of a positive rating to characterize a user
preference not only in terms of items but also in terms of classes.
2. Role of the semantics in the SBS data structure. Our segmentation
is based on a semantic data structure, which is built thanks to an ontology
and to semantic analysis tools. We validate this choice by evaluating the
difference between the number of characterizing bits both in a binary
vector built by analyzing the original words of the item descriptions and
the SBS built thanks to the semantic analysis.
3. Setting of the ϕparameter. The segmentation is built by putting
together all the users with a relevance score higher than a threshold ϕ.
This experiment sets the threshold for each class by employing the elbow
method, which evaluates the relevance score of each user for a class and
detects the point in which the score does not characterize the class any-
more, since too many users are included in the segment that represents
it.
4. Analysis of the segments. This experiment analyzes the segments of
users targeted for each class, in order to evaluate the capability of our
proposal to include also users who do not express explicit preferences for
a class but might be interested in it.
5. Performance analysis. Given a new item classified with a class, we
evaluate the number of seconds it takes to update the SBS data structure
(i.e., to perform the semantic disambiguation, evaluate the synsets in the
22
item description, and include this information in the SBS). Note that
descriptions of different lengths lead to different computational efforts, so
this analysis allows us to evaluate the performance of the approach from
different perspectives.
It should be observed that in order to validate the capability of our proposal
to detect users who are not characterized by explicit preferences for a class,
we compare with the so-called topic-based approach employed by both Google’s
AdWords and Facebook’s Core Audiences. In order to do so, in the experiments
number 4 and 5, we also build a relevance score for each user and each class, by
considering how many movies of a genre a user evaluated (i.e., we are considering
a scenario in which the topic of interest is a genre of movies, which is equivalent
to our classes). This is done since the companies did not reveal how they
associate users to topics, and in order to make a direct comparison between an
approach that uses explicit preferences and our semantic approach.
5.2. Dataset
The used dataset, i.e., Yahoo! Webscope Movie Dataset (R4), contains a
large amount of data related to user preferences expressed by the Yahoo! Movies
community that are rated on the base of two different scales, from 1 to 13 and
from 1 to 5 (we have chosen to use the latter). The training data is composed
by 7642 users (|U|), 11915 movies/items (|I|), and 211231 ratings (|R|).
The average user rating (Ru=Puru
|U|,macro-averaged) is 3.70 and the average
item rating (macro-averaged) is 3.58. The average number of ratings per user
is 27.64 and the average number of ratings per item is 17.73. All users have
rated at least 10 items and all items are rated by at least one user. The density
ratio (δ=|R|
|U|∗|I|) is 0.0023, meaning that only 0.23% of entries in the user-item
matrix are filled.
As shown in Table 4, the items are classified by Yahoo in 19 different classes
(movie genres), and it is should be noted that each item may be classified in
multiple classes.
23
01 Action/Adventure 11 Musical/Performing Arts
02 Adult Audience 12 Other
03 Animation 13 Reality
04 Art/Foreign 14 Romance
05 Comedy 15 Science Fiction/Fantasy
06 Crime/Gangster 16 Special Interest
07 Documentary 17 Suspense/Horror
08 Drama 18 Thriller
09 Kids/Family 19 Western
10 Miscellaneous
Table 4: Yahoo! Webscope R4 Genres
5.3. Metric
In order to detect the relevance score to take into account during the user
segmentation (i.e., the threshold value after which we can consider a score as
discriminant), we use the well-known elbow method. In other words, we increase
the relevance score value and calculate the variance (as shown in Equation (9),
where xdenotes the number of users involved, and nis the relevance score) of
the users involved: at the beginning we can note a low level of variance, but at
some point the level suddenly increases; following the elbow method we chose as
threshold value the number of synset occurrences used at this point.
S2=P(xi−x)
n−1(9)
5.4. Experimental Results
This section presents the results of each experiment previously presented.
5.4.1. Data Overview
In the first experiment we performed a preliminary study on the relation
between the users and the native classification of the items in the dataset, in
order to analyze the distribution of users with respect to the classes. For each
class, Figure 4 reports the number of users for which that class is the one with
most evaluations. Moreover, above each point, we indicate the ranking of the
classes, based on the number of users.
The results show that 15 out of 19 classes have more than 1000 users for
which it is the most relevant. Moreover, 6 classes are the most relevant for a
24
2 4 6 8 10 12 14 16 18 20
2,000
4,000
6,000
8,000 2
19
10
13
1
5
15
3
7
18
12
9
17
8
6
16
11
4
14
Native Classes
Number of Users
Ranking
Figure 4: User distribution for native classes
number of users between 6000 and 8000. The fact that each class is the most
relevant for a lot of users, and it does not exist a unique dominant class that is
the most relevant for all the users, ensures that the segmentation process is not
trivial (indeed, if all the users could be associated to one class, the relevance
scores for that class would be very high and the segmentation would be trivial).
In Figure 5 we see the number of items that have been classified with multiple
genres. The results show that most of the items have been classified with a single
genre and it is rare to find items classified with multiple genres (only one item
in the whole dataset has 6 co-classifications). This means that when a user
positively evaluates an item, it is possible to derive a preference also in terms of
classes, and the synsets contained in an item description characterizes the SBS
of just one class (i.e., the SBSs will not be similar, since disjoint sets of items
contribute to each binary vector).
5.4.2. Role of the Semantics in the SBS Data Structure
In order to validate our choice to represent a SBS as a semantic data struc-
ture, we built the equivalent of the SBS by considering the original words avail-
able in the item descriptions. This means that Wordnet was not employed and
25
012345678910
2
4
6
Involved items (×1000)
Number of co-classifications
Figure 5: Number of coclassification for item
W ords 63772
Synsets 91130
Dif f erence +30.02%
Table 5: Synsets and words cardinality
no synset was collected, and of course we could not perform a semantic dis-
ambiguation of the words. We did this comparison for each class and since 19
classes are involved, in order to facilitate the interpretability of the results, on
the one hand we summed the amount of 1 occurrences in the 19 SBSs, while
on the other hand we summed the amount of 1 occurrences in the 19 binary
vectors containing the words. The results presented in Table 5 show that, when
considering the words, the classes are characterized by 30% less elements, with
respect to their semantic counterpart. This shows the high relevance that the
employment of the ontology has, and how important it is to perform a semantic
disambiguation among the words. Indeed, by associating the correct seman-
tic sense to each word it is possible to avoid phenomena that characterize this
area, such as synonymity, and to have more accurate information about what
characterizes each class of items.
26
Cl ass T opic-based S BS -based Cl ass T opic-based SBS -based
1 29 1414 11 4 789
27 0 12 12 1112
34 857 13 1 47
4 9 778 14 8 1170
545 1438 15 17 1269
68 1195 16 3 270
72 287 17 15 1033
840 1369 18 16 1269
9 12 1162 19 6 535
10 1 9
Table 6: Elbow values
5.4.3. Setting of the ϕparameter
In order to set the value of ϕthat allows us to consider a class as relevant
for a user, we adopted the elbow method introduced in Section 5.3. Table 6
shows the threshold values derived from elbow method, i.e., for each class we
indicate the minimum value the relevance score of a user has to have, in order
for a user to be included in the segment of that class. In order to be able to
compare our semantic approach to a topic-based segmentation that considers
the explicitly expressed preferences, we performed this analysis for both types
of vectors that describe a class. Note that the threshold values for the SBS data
structure are much higher with respect to the topic-based values. This means
that when the semantics behind the item descriptions are considered (and not
just the explicitly expressed preferences), a user is associated to a class many
more times, thus showing the capability of our approach to capture latent links
between the users and the classes.
5.4.4. Analysis of the segments
In this section, we analyze the produced user segments. For each of the
primitive classes, we present an analysis of the segments generated by both the
baseline topic-based approach and by our SBS approach. Regarding the boolean
classes, since all the possible ways to combine multiple classes with the three
27
Class T opic-based Segments SBS S egments Shared U sers Unsh ared U sers co-classif ications %
1 208 604 206 398 394 98.99
2 0 0 0 0 0 0.00
3 177 940 147 793 786 99.12
4 53 1013 37 976 969 99.28
5 120 590 120 470 466 99.15
6 242 717 200 517 510 98.65
7 40 1518 28 1490 1482 99.46
8 117 622 117 505 499 98.81
9 99 737 92 645 639 99.07
10 0 1026 0 1026 1015 98.93
11 90 1015 77 938 931 99.25
12 87 762 75 687 682 99.27
13 0 1945 0 1945 1930 99.23
14 2 43 725 214 511 507 99.22
15 1 85 666 178 488 481 98.57
16 12 1870 9 1861 1848 99.30
17 78 818 66 752 746 99.20
18 1 96 668 193 475 468 98.53
19 22 1228 20 1208 1200 99.34
5 AND 1 82 640 82 558 552 98.9 2
5 OR 1 246 559 244 315 311 98.73
13 AND 10 0 3002 0 3002 2971 98.97
13 OR 10 0 1737 0 1737 1724 99.25
5 NOT 14 22 200 19 181 72 36.00
Table 7: Experiments result
operators are impossible to analyze, we decided to study the segments generated
through an interclass- and a superclass-based SBS by combining the two classes
with most and least items in the dataset (respectively, classes 1 and 5, and 13
and 1014); this allowed us to analyze our approach both in a scenario where a
lot of information is available and in a case in which the users expressed very
little preferences for that class.
The subclass-based segmentation was studied by considering the two classes
with which the items were most co-classified (i.e., classes 5 and 14). Table 7
presents the obtained results and the columns contain the following information:
Class contains the identifier of the class that characterizes the interest of the
14Note that class 2 is actually the class with least items, but we will show that its relevance
in the dataset is so low that it cannot be managed in practice.
28
users in it, Topic-based Segments and SBS Segments report the amount of users
included in the segment by the two approaches, Shared Users and Unshared
Users respectively report how many users have been identified by both ap-
proaches and how many have been detected with our proposal, co-classification
reports for how many unshared users a class that was relevant for them was
also co-classified with the considered class (a positive outcome means that we
added a relevant user to the segment of a class, since the class considered in
the segment is naturally correlated with a class that is relevant for the user)15 ,
and column %reports the percentage of relevant unshared users detected by
our approach (i.e., those for which a co-classified relevant class was found).
When analyzing the results of the primitive classes, we can notice that the
SBS segments contain from 3 to 155 times more users with respect to their
Topic-based counterparts. We can also notice that the difference between the
amount of users added to a segment is higher for the classes that are the relevant
for less users (i.e., classes 3, 4, 7, and 16, which in Figure 4 are all associated to
the lowest part of the figure).
In addition, we can notice that our approach is able to detect a balanced
amount of users for each class; this would allow advertisers to efficiently target
users, no matter which class is considered. A related and important character-
istic of our approach, is its capability to detect an homogeneous amount of users
regardless of how much explicit information about the preferences for the classes
is available; indeed, even the less relevant classes can lead to a targeting that
considers a high amount of users (note that for the two least relevant classes,
i.e., 10 and 13, the topic-based approach cannot detect any user, while we are
able to characterize those classes thanks to the semantics). The only exception
to this is class 2 (Adult), which is the least relevant in the dataset and the
amount of positive preferences for these items was so little that neither of the
15The only exception to this analysis regards the NOT operator, in which we analyzed how
many users had a semantic relevance score higher than the threshold in the first class but not
in the second.
29
two approaches could add users to its segment.
The very relevant classes in the dataset, such as 1 and 5, are not flooded
with too many users and elbow method has proven to be an effective criterion
to choose the threshold.
Regarding the unshared users, detected by our approach but not by the
topic-based one, we can notice that more than 98% of them are relevant, since
we found another class that is relevant for them when considering the topic-
based preferences, and whose items are co-classified with the considered class.
The analysis of the interclass-based segments (AND operator) and of the
superclass-based segments (OR operator), show very similar results to those
reported for the primitive classes. These results confirm the capability of our
approach to work well when few explicit information is available, even when the
classes are combined into a boolean one. An interesting result to analyze is the
last line of the table, related to the subclass-based segment 5 N OT 14, for which
36% of the unshared users that have been detected are relevant. When looking
for users interested by Comedy movies (class 5) that do not contain Romantic
elements (class 14), our approach detected 9 times the users of the topic-based
one; out of these 200 detected users, 72 of them (3 times the users detected by the
topic-based approach) reported a semantic relevance for class 5 but not for class
14. Regarding the remaining users, they do like both Comedy and Romance
movies, but this result shows that even if we remove the Romance elements from
the Comedy movies, a strong interest for the Comedy genre remains (in other
words, they could be targeted as users that might like Comedy movies that do
not contain Romance elements).
5.4.5. Performance Analysis
Figure 6 reports the number of seconds it takes for our approach to update
the SBS of a class once a new item receives a positive rating. Note that to
simplify the readability of the results we report just the performance of the first
100 items of the dataset. The dashed line in the figure represents the average
number of seconds considering all the values.
30
These results show that different items lead to a quite different performance.
We inspected on this result furthermore, and we saw that all the different steps
performed at the beginning of the computation, and presented in Section 4.1,
play a role in the performance of the approach. Indeed, when an item description
contains more synsets, the number of seconds necessary to complete the data
structure update is higher, but there is not a direct correlation between the
number of synsets and the performance (i.e., item 19 is not the one with the
highest number of synsets among the 100 items considered, even though it is
the one with the lowest performance). Indeed, the other steps, such as the text
preprocessing, influence the performance and lead to the different results.
Regarding the performance of the SBS update, which is the core of our
approach, it should also be noted that it lends itself well to a processing through
grid computing. Indeed, the processing of the individual items might be done
on different computers. For example, a possible optimized solution is to use
a single computer for the computation of the SBS for a subset of items, so
that the computation of the final SBS is distributed over different computers,
by employing large scale distributed computing models, like MapReduce. It
is trivial to notice that the final SBS is a combination of the output of the
individual machines through an OR operator (if a synset is relevant for an item,
it is relevant for the class).
6. Conclusions and Future Work
This paper presented a novel semantic user segmentation approach that ex-
ploits the description of the items positively evaluated by the users. The de-
tection of the segments is based on the definition of a set of binary sieves, new
entities that allow us to characterize primitive or boolean classes (i.e., set of
classes combined through boolean operations). The experimental results show
the ability of our semantic approach to model in an effective way a target of
users within the domain taken into account. Future work will test its capa-
bility to characterize clusters of users whose purchased items are semantically
31
0 20 40 60 80 100
0
5
10
15
Item ID
Time (sec)
Figure 6: Execution time
related. This will allow us to target the users in a different way, e.g., by perform-
ing group recommendations to them (i.e., by recommending items to groups of
“semantically similar” users).
Acknowledgments
This work is partially funded by Regione Sardegna under project NOMAD (Next
generation Open Mobile Apps Development), through PIA - Pacchetti Inte-
grati di Agevolazione “Industria Artigianato e Servizi” (annualit`a 2013), and
by MIUR PRIN 2010-11 under project “Security Horizons”.
References
[1] X. Gong, X. Guo, R. Zhang, X. He, A. Zhou, Search behavior based la-
tent semantic user segmentation for advertising targeting, in: Data Mining
(ICDM), 2013 IEEE 13th International Conference on, 2013, pp. 211–220.
doi:10.1109/ICDM.2013.62.
[2] S. Tu, C. Lu, Topic-based user segmentation for online advertising with latent dirichlet allocation,
in: Proceedings of the 6th International Conference on Advanced Data
32
Mining and Applications - Volume Part II, ADMA’10, Springer-Verlag,
Berlin, Heidelberg, 2010, pp. 259–269.
URL http://dl.acm.org/citation.cfm?id=1948448.1948476
[3] A. Spink, B. J. Jansen, D. Wolfram, T. Saracevic,
From e-sex to e-commerce: Web search changes, Computer 35 (3) (2002)
107–109. doi:10.1109/2.989940.
URL http://dx.doi.org/10.1109/2.989940
[4] S. Y. Rieh, H. I. Xie, Analysis of multiple query reformulations on the web: The interactive information retriev
Inf. Process. Manage. 42 (3) (2006) 751–768.
doi:10.1016/j.ipm.2005.05.005.
URL http://dx.doi.org/10.1016/j.ipm.2005.05.005
[5] P. Boldi, F. Bonchi, C. Castillo, S. Vigna,
From ”dango” to ”japanese cakes”: Query reformulation models and patterns,
in: Proceedings of the 2009 IEEE/WIC/ACM International Joint Confer-
ence on Web Intelligence and Intelligent Agent Technology - Volume 01,
WI-IAT ’09, IEEE Computer Society, Washington, DC, USA, 2009, pp.
183–190. doi:10.1109/WI-IAT.2009.34.
URL http://dx.doi.org/10.1109/WI-IAT.2009.34
[6] G. Armano, A. Giuliani, E. Vargiu, Semantic enrichment of contextual
advertising by using concepts, in: J. Filipe, A. L. N. Fred (Eds.), KDIR
2011 - Proceedings of the International Conference on Knowledge Discovery
and Information Retrieval, Paris, France, 26-29 October, 2011, SciTePress,
2011, pp. 232–237.
[7] G. Armano, A. Giuliani, E. Vargiu,
Studying the impact of text summarization on contextual advertising,
in: F. Morvan, A. M. Tjoa, R. Wagner (Eds.), 2011 Database and Expert
Systems Applications, DEXA, International Workshops, Toulouse, France,
August 29 - Sept. 2, 2011, IEEE Computer Society, 2011, pp. 172–176.
URL http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=6059238
33
[8] R. Saia, L. Boratto, S. Carta, Semantic coherence-based user profile mod-
eling in the recommender systems context, in: Proceedings of the 6th In-
ternational Conference on Knowledge Discovery and Information Retrieval,
KDIR 2014, Rome, Italy, October 21-24, 2014, SciTePress, 2014, pp. 154–
161.
[9] R. D. Burke, M. Ramezani, Matching recommendation technologies and domains,
in: F. Ricci, L. Rokach, B. Shapira, P. B. Kantor (Eds.), Recommender
Systems Handbook, Springer, 2011, pp. 367–386.
URL http://www.springerlink.com/content/978-0-387-85819-7
[10] P. Lops, M. de Gemmis, G. Semeraro, Content-based recommender sys-
tems: State of the art and trends, in: F. Ricci, L. Rokach, B. Shapira,
P. B. Kantor (Eds.), Recommender Systems Handbook, Springer, 2011,
pp. 73–105.
[11] C. Gustav Johannsen, Understanding users: from man-made typologies to
computer-generated clusters, New Library World 115 (9/10) (2014) 412–
425. doi:10.1108/NLW-05-2014-0052.
[12] S. C. Bourassa, F. Hamelink, M. Hoesli, B. D. MacGregor,
Defining housing submarkets, Journal of Housing Economics 8 (2)
(1999) 160 – 183. doi:http://dx.doi.org/10.1006/jhec.1999.0246.
URL http://www.sciencedirect.com/science/article/pii/S1051137799902462
[13] A. Nairn, P. Bottomley, Something approaching science? cluster analysis procedures in the crm era,
in: H. E. Spotts (Ed.), Proceedings of the 2002 Academy of Marketing
Science (AMS) Annual Conference, Developments in Marketing Science:
Proceedings of the Academy of Marketing Science, Springer International
Publishing, 2003, pp. 120–120. doi:10.1007/978-3-319-11882-6_40.
URL http://dx.doi.org/10.1007/978-3-319-11882-6_40
[14] S. Dolnicar, K. Lazarevski, Methodological reasons for the theory/practice
divide in market segmentation, Journal of Marketing Management 25 (3-4)
(2009) 357–373. doi:10.1362/026725709X429791.
34
[15] S. Dibb, L. Simkin, A program for implementing market segmenta-
tion, Journal of Business & Industrial Marketing 12 (1) (1997) 51–65.
doi:10.1108/08858629710157931.
[16] H. Zhuge, Semantic linking through spaces for cyber-physical-socio intelligence: A methodology,
Artif. Intell. 175 (5-6) (2011) 988–1019.
doi:10.1016/j.artint.2010.09.009.
URL http://dx.doi.org/10.1016/j.artint.2010.09.009
[17] H. Zhuge, Interactive semantics, Artif. Intell. 174 (2) (2010) 190–204.
doi:10.1016/j.artint.2009.11.014.
URL http://dx.doi.org/10.1016/j.artint.2009.11.014
[18] J. Yan, N. Liu, G. Wang, W. Zhang, Y. Jiang, Z. Chen,
How much can behavioral targeting help online advertising?, in: Proceed-
ings of the 18th International Conference on World Wide Web,
WWW ’09, ACM, New York, NY, USA, 2009, pp. 261–270.
doi:10.1145/1526709.1526745.
URL http://doi.acm.org/10.1145/1526709.1526745
[19] H. Beales, The value of behavioral targeting, Network Advertising Initia-
tive.
[20] Y. Chen, D. Pavlov, J. F. Canny, Large-scale behavioral targeting, in: Pro-
ceedings of the 15th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD ’09, ACM, New York, NY, USA, 2009,
pp. 209–218. doi:10.1145/1557019.1557048.
URL http://doi.acm.org/10.1145/1557019.1557048
[21] G. Armano, E. Vargiu, A unifying view of contextual advertising and rec-
ommender systems, in: A. L. N. Fred, J. Filipe (Eds.), KDIR 2010 - Pro-
ceedings of the International Conference on Knowledge Discovery and Infor-
mation Retrieval, Valencia, Spain, October 25-28, 2010, SciTePress, 2010,
pp. 463–466.
35
[22] A. Addis, G. Armano, A. Giuliani, E. Vargiu, A recommender system based
on a generic contextual advertising approach, in: Proceedings of the 15th
IEEE Symposium on Computers and Communications, ISCC 2010, Ric-
cione, Italy, June 22-25, 2010, IEEE, 2010, pp. 859–861.
[23] E. Vargiu, A. Giuliani, G. Armano,
Improving contextual advertising by adopting collaborative filtering,
ACM Trans. Web 7 (3) (2013) 13:1–13:22.
doi:10.1145/2516633.2516635.
URL http://doi.acm.org/10.1145/2516633.2516635
[24] J. Mazanee, Market segmentation, in: Encyclopedia of Tourism, London:
Routledge, 2000.
[25] S. Dolniˇcar, Beyond “commonsense segmentation”: A systematics of seg-
mentation approaches in tourism, Journal of Travel Research 42 (3) (2004)
244–250.
[26] J. H. Myers, E. M. Tauber, Market Structure Analysis, American Market-
ing Association, 1977.
[27] M. Wedel, W. A. Kamakura, Market Segmentation: Conceptual and
Methodological Foundations (International Series in Quantitative Market-
ing), Kluwer Academic Publishers, 2000.
[28] J. Bian, A. Dong, X. He, S. Reddy, Y. Chang,
User action interpretation for online content optimization, IEEE
Trans. on Knowl. and Data Eng. 25 (9) (2013) 2161–2174.
doi:10.1109/TKDE.2012.130.
URL http://dx.doi.org/10.1109/TKDE.2012.130
[29] Z. Yao, T. Eklund, B. Back, Using som-ward clustering and predictive analytics for conducting customer segmen
in: Proceedings of the 2010 IEEE International Conference on Data Mining
Workshops, ICDMW ’10, IEEE Computer Society, Washington, DC, USA,
36
2010, pp. 639–646. doi:10.1109/ICDMW.2010.121.
URL http://dx.doi.org/10.1109/ICDMW.2010.121
[30] Y. K. Zhou, B. Mobasher, Web user segmentation based on a mixture of factor analyzers,
in: Proceedings of the 7th International Conference on E-Commerce and
Web Technologies, EC-Web’06, Springer-Verlag, Berlin, Heidelberg, 2006,
pp. 11–20. doi:10.1007/11823865_2.
URL http://dx.doi.org/10.1007/11823865_2
[31] X. Wu, J. Yan, N. Liu, S. Yan, Y. Chen, Z. Chen,
Probabilistic latent semantic user segmentation for behavioral targeted advertising,
in: Proceedings of the Third International Workshop on Data Mining and
Audience Intelligence for Advertising, ADKDD ’09, ACM, New York, NY,
USA, 2009, pp. 10–17. doi:10.1145/1592748.1592751.
URL http://doi.acm.org/10.1145/1592748.1592751
[32] S. A. Munson, P. Resnick, Presenting diverse political opinions: How and how much,
in: Proceedings of the SIGCHI Conference on Human Factors in Comput-
ing Systems, CHI ’10, ACM, New York, NY, USA, 2010, pp. 1457–1466.
doi:10.1145/1753326.1753543.
URL http://doi.acm.org/10.1145/1753326.1753543
[33] E. Pariser, The Filter Bubble: What the Internet Is Hiding from You,
Penguin Group , The, 2011.
[34] L. Festinger, A theory of cognitive dissonance, Vol. 2, Stanford university
press, 1962.
[35] S. Park, S. Kang, S. Chung, J. Song,
Newscube: delivering multiple aspects of news to mitigate media bias,
in: D. R. O. Jr., R. B. Arthur, K. Hinckley, M. R. Morris, S. E. Hudson,
S. Greenberg (Eds.), Proceedings of the 27th International Conference on
Human Factors in Computing Systems, CHI 2009, Boston, MA, USA, April
4-9, 2009, ACM, 2009, pp. 443–452. doi:10.1145/1518701.1518772.
URL http://doi.acm.org/10.1145/1518701.1518772
37
[36] G. Salton, C. Buckley, Term-weighting approaches in automatic text retrieval,
Inf. Process. Manage. 24 (5) (1988) 513–523.
doi:10.1016/0306-4573(88)90021-0.
URL http://dx.doi.org/10.1016/0306-4573(88)90021-0
[37] H. Zhuge, Y. Sun, The schema theory for semantic link network,
Future Generation Comp. Syst. (3) (2010) 408–420.
doi:10.1016/j.future.2009.08.012.
URL http://dx.doi.org/10.1016/j.future.2009.08.012
[38] A. Addis, G. Armano, E. Vargiu, Assessing progressive filtering to perform
hierarchical text categorization in presence of input imbalance, in: A. L. N.
Fred, J. Filipe (Eds.), KDIR 2010 - Proceedings of the International Confer-
ence on Knowledge Discovery and Information Retrieval, Valencia, Spain,
October 25-28, 2010, SciTePress, 2010, pp. 14–23.
[39] C. Fellbaum, WordNet: An Electronic Lexical Database, Bradford Books,
1998.
[40] K. Toutanova, D. Klein, C. D. Manning, Y. Singer,
Feature-rich part-of-speech tagging with a cyclic dependency network,
in: Proceedings of the 2003 Conference of the North American
Chapter of the Association for Computational Linguistics on Human
Language Technology - Volume 1, NAACL ’03, Association for Com-
putational Linguistics, Stroudsburg, PA, USA, 2003, pp. 173–180.
doi:10.3115/1073445.1073478.
URL http://dx.doi.org/10.3115/1073445.1073478
[41] G. Salton, A. Wong, C. S. Yang, A vector space model for automatic in-
dexing, Commun. ACM 18 (11) (1975) 613–620.
38
