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.
Noname manuscript No.
(will be inserted by the editor)
A Semantic Approach to Remove Incoherent Items
From a User Profile and Improve the Accuracy of a
Recommender System
Roberto Saia ·Ludovico Boratto ·
Salvatore Carta
Received: 29 July 2015 / Accepted: 31 March 2016
Abstract Recommender systems usually suggest items by exploiting all the
previous interactions of the users with a system (e.g., in order to decide the
movies to recommend to a user, all the movies she previously purchased are
considered). This canonical approach sometimes could lead to wrong results
due to several factors, such as a change in user preferences over time, or the
use of her account by third parties. This kind of incoherence in the user pro-
files defines a lower bound on the error the recommender systems may achieve
when they generate suggestions for a user, an aspect known in literature as
magic barrier. This paper proposes a novel dynamic coherence-based approach
to define the user profile used in the recommendation process. The main aim
is to identify and remove, from the previously evaluated items, those not se-
mantically adherent to the others, in order to make a user profile as close as
possible to the user’s real preferences, solving the aforementioned problems.
Moreover, reshaping the user profile in such a way leads to great advantages
in terms of computational complexity, since the number of items considered
during the recommendation process is highly reduced. The performed experi-
ments show the effectiveness of our approach to remove the incoherent items
from a user profile, increasing the recommendation accuracy.
Keywords User Profiling ·Semantic Analysis ·Magic Barrier ·Accuracy
This work is partially funded by Regione Sardegna under pro ject NOMAD (Next generation
Open Mobile Apps Development), through PIA - Pacchetti Integrati di Agevolazione “In-
dustria Artigianato e Servizi” (annualit`a 2013), and by MIUR PRIN 2010-11 under pro ject
“Security Horizons”.
Dipartimento di Matematica e Informatica - Universit`a di Cagliari
Via Ospedale 72 - 09124 Cagliari, Italy
E-mail: {roberto.saia, ludovico.boratto, salvatore}@unica.it
2 Roberto Saia et al.
1 Introduction
The rapid growth of the number of companies that sell goods through the Word
Wide Web generates an enormous amount of valuable information, which can
be exploited to improve the quality and efficiency of the sales criteria [48]. Be-
cause of the widely-known information overload problem, it became necessary
to deal with the large amounts of data available on the Web [57]. Recommender
Systems (RS) [42] represent an effective response to this problem, by filtering
the huge amount of information about the customers in order to get useful
elements to produce suggestions to them [7,1, 56]. The denomination RS de-
notes a set of software tools and techniques providing to a user suggestions for
items, where the term item is used to indicate what the system recommends to
the users. In this paper we address one of the most important aspects related
to the recommender systems, i.e., how to represent a user profile, so that it
only contains accurate information about a user, allowing a system to generate
effective recommendations.
The main motivation behind this work is that most of the solutions re-
garding the user profiling in the recommender systems context involve the
interpretation of the whole set of items previously evaluated by a user, in
order to measure their similarity with those that she did not consider yet,
and recommend the most similar ones. Indeed, the recommendation process
is usually based on the principle that user preferences remain unchanged over
time and this can be true in many cases, but it is not the norm due to the
existence of temporal dynamics in their preferences [26, 28, 58]. Therefore, a
static approach to user profiling can lead toward wrong results due to various
factors, such as a simple change of preferences over time or the temporary use
of their own account by other people.
Several works [3,21] have shown that some user ratings might be considered
as outliers, due to the fact that the same user may rate the same item with
different ratings, at different moments in time. This is a well-known problem,
which in the literature is defined as magic barrier [10,20,44], a term used
to identify the point where, due to the noise in the data, the performance
and accuracy of an algorithm cannot be further improved. After the magic
barrier has been reached, any improvement in terms of accuracy might mean
an overfitting instead of a performance enhancement. Thus, the primary aim of
the approach we introduce is to measure the similarity between a single item
and the others in the user profile, in order to improve the recommendation
process by discarding the items that are highly dissimilar with the others in
the user profile. In the literature, the coherence of an item with respect to
the user profile is usually measured as the variance in the feature space that
defines the items, typically based on the ratings given by the users [10]. This is
done by employing several metrics, such as the entropy, the mean value, or the
standard deviation. Differently from the approaches at the state of the art, in
this paper we consider the semantic distance between the concepts expressed
by each item in a user profile, and the concepts expressed by the other ones.
A Semantic Approach to Remove Incoherent Items From a User Profile 3
This way to proceed presents a twofold advantage: firstly, it allows us to
evaluate the coherence of an item in a more extensive way (i.e., by employ-
ing semantic concepts) w.r.t. a limited mathematical approach; secondly, it
reduces the cause of the magic barrier problem. This happens because the
assumption of the magic barrier problem is the presence of incoherent items
in the user profiles. Considering that our approach removes them, keeping in
the user profiles only those items that are coherent with each other, we can
consider any observed improvement as real, instead that a mere side effect
(i.e., an overfitting).
To perform the task of removing semantically incoherent items from a user
profile, we introduce the Dynamic Coherence-Based Modeling (DCBM), an al-
gorithm based on the concept of Minimum Global Coherence (MGC), a metric
that allows us to measure the semantic similarity between a single item and
the others within the user profile. Moreover, the algorithm takes into account
two factors, i.e., the position of each item in the chronology of the user choices,
and the distance from the mean value of the global similarity (as “global” we
mean all the items in a user profile). These metrics allow us to remove in a
selective way any item that could make the user profiles non-adherent to the
real preferences of the users. The main idea is that the more information in the
user profile is coherent, the more the recommendations based on this profile
will be reliable. Through our approach, the evaluation process of the items
coherence has been moved from a domain based on rigorous mathematical cri-
teria (i.e., variance of the user ratings in the feature space), to a new semantic
domain, which presents a considerable advantage in terms of evaluation flex-
ibility. An important aspect related with the proposed approach is its ability
to operate in any domain characterized by a textual description of the items,
even in absence of a user rating about them (e.g., implicit user preferences
collected from the users’ browsing sessions).
In order to validate the capability of our approach to produce accurate
user profiles, we are going to include the DCBM algorithm into state-of-the-art
recommender systems (i.e., SVD [25] and a User-Based Collaborative Filtering
approach [35]) and evaluate the accuracy of the performed recommendations.
Since the task of a recommender system that predicts the interest of the users
for the items relies on the information included in a user profile, more accurate
user profiles lead to an improved accuracy of the whole recommender system.
Experimental results show its capability to remove the incoherent items from
the user profiles, increasing the accuracy of the recommendations and reducing
the computational complexity of the system, also in the context of the non-
semantic approaches as those taken into account (i.e., SVD and User-Based
Collaborative Filtering approaches).
The contributions of the paper are summarized as follows:
–we introduce a novel algorithm to remove incoherent items from a user
profile, with the aim to improve the recommendation accuracy;
–we integrate our algorithm into state-of-the-art recommender systems, in
order to improve their effectiveness and validate our proposal;
4 Roberto Saia et al.
–we performed experiments on two real-world datasets and two state-of-the-
art recommender systems, which compare the accuracy of a recommender
system before and after the use of our approach.
This paper is based on the work presented in [43], which was completely
rewritten and extended in the following ways: (i) we provide a formal notation
and definition of the problem we tackle in this work, (ii) we extend the pre-
sentation of the background concepts behind this proposal, (iii) we provide
more details on the proposed approach, (iv) we compare our proposal against
two state-of-the-art recommender systems and two datasets.
The rest of the paper is organized as follows: Section 2 presents related work
on user profiling; in Section 3 we introduce the background on the concepts and
the problem handled by our proposal; Section 4 contains the formal definition
of our problem, the details of the DCBM algorithm, and its integration into a
recommender system; Section 5 presents the experimental framework used to
evaluate our approach; Section 6 contains conclusions and future work.
2 Related Work
In order to lead the potential buyers toward a number of well-targeted sugges-
tions, related to the large amount of goods or services available today through
the electronic commerce circuits, a Recommender System plays a determinant
role, since it is able to investigate on the user preferences, suggesting them
only potentially interesting items. In order to identify them, a recommender
system has to predict that an item is worth recommending [42].
The first class of systems that was developed employs the so-called Col-
laborative Filtering approach [21,23,31, 47], which is based on the assumption
that users have similar preferences on a item, if they already have rated other
similar items [54]. Content-based recommender systems, instead, suggest to
users items whose content is similar to that of the items they previously eval-
uated [33,38]. The early systems used relatively simple retrieval models, such
as the Vector Space Model, with the basic TF-IDF weighting (which is pre-
sented in detail in the next section). Examples of systems that employ this
type of content filtering are [9,11, 29,37]. Due to the fact that the approach
based on a simple bag of words is not able to perform a semantic disambigua-
tion of the words in an item description, content-based recommender systems
evolved and started employing external sources of knowledge (e.g., ontologies)
and semantic analysis tools, to improve their accuracy [13, 14].
When producing personalized recommendations to users, the first require-
ment is to understand the needs of the users and, according to them, to build
a user profile that models these needs. There are several approaches to cre-
ate user profiles: some of them focus on short-term user profiles that capture
features of the user’s current search context [12,18,50], while others accommo-
date long-term profiles that capture the user preferences over a long period of
time [8,15, 34]. As shown in [58], compared with the short-term user profiles,
the use of a long-term user profile generally produces more reliable results,
A Semantic Approach to Remove Incoherent Items From a User Profile 5
at least when the user preferences are fairly stable over a long time period.
Otherwise, we need a specific strategy able to manage the changes in the user
profile that do not reflect the preferences of the user and that represent a form
of “noise”.
The most common strategies to get useful information to build the user
profiles are two, i.e., explicit or implicit. Explicit profiling strategies directly
request to the users different forms of preference information, from categori-
cal preferences [15, 34] to simple result ratings [8]. Instead, implicit profiling
strategies attempt to infer this information by analyzing the users’ behavior,
and without a direct interaction with them while they perform actions in a
website [15,32,40].
However, the strategy usually adopted is the implicit one, where the user
preferences are inferred without a direct interaction with her. This is the com-
mon approach because the explicit strategy presents some problems, such as
those related with the privacy aspects (many users do not like to reveal infor-
mation about their preferences), and those related with the form filling process
(many users do not like to spend their time for this activity, and it is proved
that the accuracy of the information depends on the time needed to provide
them).
The implicit approach usually requires long-term user profiles, where the
information about the preferences is considered over an extended period of
time. However, there are some implicit approaches that involve a short-term
profiling, related to the particular context in which the system operates [50].
Regardless of the type of profiling that is adopted (e.g., long-term or short-
term), there is a common problem that may affect the goodness of the obtained
results, i.e., the capability of the information stored in the user profile to
lead toward reliable recommendations. In order to face the problem of dealing
with unreliable information in a user profile, the state of art proposes dif-
ferent strategies. Several approaches, such as [26], take advantage from the
Bayesian analysis of the user provided relevance feedback, in order to detect
non-stationary user interests. Also exploiting the feedback information pro-
vided by the users, other approaches such as [58] make use of a tree-descriptor
model to detect shifts in user interests. Another technique exploits the knowl-
edge captured in an ontology [49] to obtain the same result, but in this case
it is necessary that the users express their preferences about items through an
explicit rating.
There are also other different strategies that try to improve the accuracy
of the information in the profiles by collecting the implicit feedbacks of the
users during their natural interactions with the system (reading-time, saving,
etc.) [24]. However, it should be pointed out that most of the strategies used
in this area are effective only in specific contexts, such as for instance [60],
where a novel approach to automatically model the user profile, according
to the change of her preferences, is designed for the articles recommendation
context.
With regard to the analysis of information related to user profiles and
items, there are several ways to operate and most of them work by using
6 Roberto Saia et al.
the bag-of-words model, an approach where the words are processed without
taking into account the correlation between the terms [26,58]. This trivial way
to manage the information usually does not lead toward good results, and
just for this reason there are some more sophisticated alternatives, such as
the semantic analysis of the content in order to model the preferences of a
user [39]. In [51,52,53], the problem of modeling semantically correlated items
was tackled, but the authors consider a temporal correlation and not the one
between the items and a user profile.
It should be noted that there is a common issue that afflicts the recom-
mendation approaches, related to the concept of item incoherence. This is a
problem that in the literature is identified as magic barrier [20], a term used
to define the theoretical boundary for the level of optimization that can be
achieved by a recommendation algorithm on transactional data [45]. The eval-
uation models assume as a ground truth that the transactions made in the past
by the users, and stored in their profiles, are free of noise. This is a concept
that has been explored in [4, 3], where a study aimed to capture the noise in a
service that operates in a synthetic environment was performed. It should be
noted that this is an aspect that, in the context of the recommender systems,
was mentioned for the first time in 1995, in a work aimed at discussing the
concept of reliability of users in terms of rating coherence [21].
Our approach differs from the others in the literature, since it does not need
to focus on a specific type of profile (i.e., short-term or long-term), but it can
operate with any type of data that contains a textual description, overcom-
ing the limitation introduced by the magic barrier from a novel perspective,
represented by the semantic analysis of the items.
3 Background
In this section we provide some details concerning two key concepts involved
in the context of this work, i.e., the spatial representation of a text docu-
ment based on the Vector Space Model, and the functionalities offered by the
WordNet environment.
3.1 Vector Space Model
Many content-based recommender systems use relatively simple retrieval mod-
els [33], such as the Vector Space Model (VSM), with the basic TF-IDF weight-
ing. VSM is a spatial representation of text documents, where each document
is represented by a vector in a n-dimensional space, and each dimension is
related to a term from the overall vocabulary of a specific document col-
lection. In other words, every document is represented as a vector of term
weights, where the weight indicates the degree of association between the doc-
ument and the term. Let D={d1, d2, ..., dN}indicate a set of documents, and
dj={t1, t2, ..., tN}, t ∈Tbe the set of terms in a document. The dictionary
A Semantic Approach to Remove Incoherent Items From a User Profile 7
Tis obtained by applying some standard Natural Language Processing (NLP)
operations, such as tokenization, stop-words removal and stemming, and ev-
ery document djis represented as a vector in a n-dimensional vector space,
so dj={w1j, w2j, ..., wnj }where wnj represents the weight for term tnin
document dj. The major problems during the document representation with
the VSM are the weighting of the terms and the evaluation of the similarity
of the vectors. The most commonly used way to estimate the term weighting
is based on TF-IDF, a trivial approach that uses empirical observations of the
documents’ terms [46].
3.2 WordNet Environment
Due to the fact that an approach based on a simple bag of words is not able to
perform a semantic disambiguation of the words in an item description, also
motivated by the fact that exploiting a taxonomy for categorization purposes
is an approach recognized in the literature [2], and by the fact that a semantic
analysis is useful to improve the accuracy of a classification [5,6], in order
to perform the similarity measures used in this work we decided to exploit
the functionalities offered by the WordNet environment. WordNet is a large
lexical database of 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 155,287 words, organized into
117,659 synsets for a total of 206,941 word-sense pairs [17].
3.2.1 WordNet Structure
The main relation among words in WordNet is the synonymy and, in order
to represent these relations, the dictionary is based on synsets, i.e., 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 num-
ber of conceptual relations. Word forms with several distinct meanings are
represented by as many distinct synsets, so that each form-meaning pair in
WordNet is 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 ad-
verbs. Both nouns and verbs are organized into precise hierarchies, defined by
hypernym or is-a relationships. 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)
8 Roberto Saia et al.
(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 electro-
magnetic waves)
2. POS=verb
(a) radio (transmit messages via radio waves)
Each synset has a unique index and shares its properties, such as a gloss
or dictionary definition.
In the case of nouns and verbs (the organization of adjectives and adverbs
is slightly different) the WordNet hierarchies are organized into several base
types (25 primitive groups for the nouns and 15 for the verbs), and all primitive
groups ultimately go up to an abstract root node. As we can imagine, the
network of nouns is far deeper than that of the other parts-of-speech. The verbs
instead present a more bushy structure, and the adjectives are distributed into
many clusters, as well as the adverbs, since these last are defined in terms
of the adjectives (i.e., they are derived from adjectives and thus inherit the
structure from them). Due to the similarity measure chosen for our work, we
consider only the nouns and the verbs, exploiting the state-of-art semantic-
based approach to item recommendation based on the WordNet synsets [39], to
evaluate the semantic similarity between the items stored in the user profiles.
4 Item Recommendation with Incoherent Items Removal
The definition of the problem handled by our proposal, the notation used in
the problem statement, the details about the implementation of the proposed
algorithm, and its integration on a recommender system, are described in the
following (Sections 4.1 and 4.2).
4.1 Problem Definition
Here, after introducing the adopted notation, we define the problem handled
by our proposal.
Definition 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 possi-
ble 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, v= 3).
Definition 2 (User items) Given the set of positive preferences P+, we
denote as I+={i∈I| ∃(u, i, v)∈P+}the set of items for which there is a
A Semantic Approach to Remove Incoherent Items From a User Profile 9
positive preference, and as Iu={i∈I| ∃(u, i, v)∈P+∧u∈U}the set of
items a user ulikes.
Definition 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 vec-
tor 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 term 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 4.2.1.
Definition 4 (Semantic user model) Given the set of positively evaluated
items by a user Iu, we define a semantic user model Muas the set of synsets
in the semantic descriptions of the items in Iu. More formally, Mu={sw|
sw∈sdm∧im∈Iu,∀im∈Iu}.
Definition 5 (Item coherence) An item i∈Iuis coherent with the rest
of the items in the user profile Iu, if the similarity between the semantic
description sdiof the item and the union of the semantic descriptions of the
rest of the items (i.e., Mu\sdi) is higher than a threshold value.
Problem 1 Given a set of items Iuthat a user likes, our objective is to extract
a set Iu⊆Iu, such that each item i∈Iuis coherent with the others.
4.2 Our Approach
As already highlighted during the description of the limits that affect the
user profiling activity, individual profiles need to be as adherent as possible to
the real preferences of the users, because they are exploited to predict their
future interests. For this reason, in this section we propose a novel approach
defined Dynamic Coherence-Based Modeling (DCBM) that allows us to find
and remove the incoherent items from user profiles, regardless of the chosen
profiling method. The implementation on a recommender system of the DCBM
is described in the following subsections.
4.2.1 Data Preprocessing
Before comparing the similarity between the items in a user profile, we need to
follow several preprocessing steps. The first step 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 [55]. In the second step,
we remove punctuation marks and stop words, i.e., the insignificant words
(such as adjectives, conjunctions, etc.) that represent noise in the semantic
analysis. Several stop-words lists can be found on the Internet, and in this
10 Roberto Saia et al.
work we have used a list of 429 stop words made available with the Onix Text
Retrieval Toolkit1. In the last step, after we have determined the lemma of each
word using the Java API implementation for WordNet Searching JAWS2, we
perform the so-called word sense disambiguation, a process where the correct
sense of each word is determined, which permits us to evaluate the semantic
similarity in a precise way. The best sense of each word in a sentence (i.e.,
the selection of the real meaning of a word in the context where it is used)
was found through the Java implementation of the adapted Lesk algorithm
provided by the Denmark Technical University similarity application [46]. All
the collected synsets form the set S={s1,...,sW}defined in Section 4.1. 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 sdi; each element of the
vector sdi[w] is 1 if the corresponding synset appears in the item description,
and 0 otherwise.
4.2.2 Semantic Similarity
The most used semantic similarity measures are five, and are those defined by
Leacock and Chodorow [27], Jiang and Conrath [22], Resnik [41], Lin [30], and
Wu and Palmer [59]. Each of them evaluates the semantic similarity between
two WordNet synsets, and we calculate the semantic similarity by using Wu
and Palmer’s measure, a method based on the path length between a pair of
concepts (WordNet synsets), which in the literature is considered to be the
most accurate when generating the similarities [16, 13].
Given a set Xof iWordNet synsets x1, x2, ..., xithat are related to an item
description, and a set Yof jWordNet synsets y1, y2, ..., yjrelated to another
item description, a set Q, which contains all the possible pairs between the
synsets in the set Xand the synsets in the set Y, is defined as in Equation 1.
Q= (hx1, y1i,hx1, y2i,...,hxi, yji)∀x∈X, y ∈Y(1)
In the next step, a subset Zof the pairs in Q(i.e., Z⊆Q) that have at
least an element with the same POS is created (Equation 2).
Z={hxi, yji | P OS (xi) = P OS(yj)}(2)
The metric measures the similarity between concepts in an ontology (in
our case it is WordNet), as shown in Equation 3.
simW P (x, y) = 2·A
B+C+ (2 ·A)(3)
Assuming that the Least Common Subsumer (LCS) of two concepts x
and yis the most specific concept that is an ancestor of both x and y, where
the concept tree is defined by the is-a relation, in Equation 3 we have that
1http://www.lextek.com/manuals/onix/stopwords.html
2http://lyle.smu.edu/ tspell/jaws/index.html
A Semantic Approach to Remove Incoherent Items From a User Profile 11
A=depth(LCS(x,y)),B=length(x,LCS(x,y)),C=length(y,LCS(x,y)). We can
note that B+Crepresents the path length from xand y, while Aindicates
the global depth of the path in the taxonomy.
In the example of Fig. 1, v4is the parent (and also ancestor) of v5, while
v2is an ancestor of both v5and v3. In this case, the LCS of v3and v5is v2,
since it is the most specific concept that is an ancestor of both v3and v5. Note
that while v1is a common subsumer of both v3and v5, it is not the least,
since there is still a child of v1(in this case it is v2), which is also a common
subsumer of both v5and v3.v4is not the least common subsumer since it is
not an ancestor of v3.
v1
. . .
. . .. . .
v2
v3
v4
...
v5
Fig. 1: WordNet Relationships Tree
In order to calculate the Wu and Palmer similarity between v3and v5, we
first determine that the least common subsumer of v3and v5is v2. Next, we
determine that the length of the path from v3to v2is 1, that the length of
the path from v5to v2is 2, and that the depth of v2is 1 (i.e., the distance
from v2to the root vertex v1). Now we can determine the similarity between
the synsets v3and v5, as shown in Equation 4.
simW P (v3, v5) = 2·1
2 + 1 + (2 ·1)= 0.40 (4)
The similarity between two items is defined as the sum of the similarity
score of all the item pairs, divided by its cardinality (the subset Zof WordNet
synsets with a common part-of-speech), as shown in Equation 5.
simW P (X , Y ) = P
(x,y)∈Z
simW P (x, y)
|Z|(5)
This similarity metric is employed by our algorithm to compute the coher-
ence of an item with the rest of the semantic user profile.
12 Roberto Saia et al.
4.2.3 Dynamic Coherence-Based Modeling
For the purpose of being able to make effective recommendations to users,
their profiles need to store only the descriptions of the items that really reflect
their preferences.
In order to identify which items positively evaluated by a user (i∈Iu) do
not reflect her preference, representing for instance the result of past wrong
choices or the use by third parties of her account, the Dynamic Coherence-
Based Modeling (DCBM) algorithm measures the Minimum Global Coherence
(MGC) of each single item description with the set of the other items present
in her profile. In other words, through MGC, the most dissimilar item with
respect to the other items is identified.
The Wu and Palmer similarity metric previously presented can be used
to calculate the M GC, as shown in Equation 6 (sdidenotes the semantic
description of an item i, and Mu\sdiindicates the semantic user model from
which the synsets in sdihave been removed).
MGC = argmin
i∈IusimW P (sdi, Mu\sdi)(6)
The basic idea is to isolate each individual item iin a user profile, seman-
tically described by sdi, and then measure its similarity with respect to the
other items (i.e., the merging of the synsets of the rest of the items), in order
to obtain a measure of its coherence within the overall context of the entire
profile.
In other words, in order to identify the most distant element from the
general context of the evaluated items, we are exploiting a basic principle
of the differential calculus, because the MGC value shown upon is nothing
other than the maximum negative slope, which is calculated by finding the
ratio between the changing on yaxis and the changing on xaxis. This is
demonstrated in Theorem 1.
Theorem 1 The Minimum Global Coherence coefficient corresponds to the
maximum negative slope.
Proof Placing on the xaxis the user iterations in a chronological order, and
on the yaxis the corresponding values of GS (Global Similarity) calculated
as simW P (sdi, Mu\sdi),∀i∈Iu, we can trivially calculate the slope value
(denoted by the letter m), as shown in Equation 7.
m=△y
△x=f(x+△x)−f(x)
△x(7)
The mathematics of differential calculus defines the slope of a curve at a
point as the slope of the tangent line at that point. Since we are working with a
series of points, the slope may be calculated not at a single point but between
two points. Considering that for each current user iteration △xis always equal
to 1 (in fact, for Nuser iterations we have that 1 −0 = 1, 2 −1 = 1, ...,
A Semantic Approach to Remove Incoherent Items From a User Profile 13
1 2 3 4 5 6 7 8 9 10 11
0.26
0.28
0.3
0.32
0.34
R1R2R3
MGC
GS
x(UserI terations)
y(GS)
Fig. 2: The maximum negative slope corresponds to the value of M GC
x y m
1 0.2884 +0.2884
2 0.2967 +0.0083
3 0.2772 -0.0195
4 0.3202 +0.0430
5 0.2724 -0.0478
6 0.2886 +0.0162
7 0.2708 -0.0178
8 0.3066 +0.0358
9 0.3188 +0.0122
10 0.2691 -0.0497
11 0.2878 +0.0187
Table 1: User profile sample data
N−(N−1) = 1), the slope value mwill always be equal to f(x+△x)−f(x).
As Equation 8 shows, where simW P (Iu) denotes simW P (sdi, Mu\sdi),∀i∈Iu,
the maximum negative slope corresponds to the value of M GC .
min △y
△x=min simW P (Iu)
1=MGC (8)
In Fig. 2, which displays the data reported in Table 1, we can see what we
just said in a graphical way.
In order to avoid the removal of an item that might correspond to a recent
change in the preferences of the user or an item not semantically distant enough
from the context of the remaining items, the DCBM algorithm considers an
item as incoherent and removes it, only if it meets the following conditions:
1. it is located in the first part of the user iteration history. Based on this
first requirement, an item is considered far from the user’s preferences only
when it goes up in the first part of the iterations. This condition is checked
thanks to a parameter r, taken as input by the algorithm, which defines
the removal area, i.e., the percentage of a user profile where an item can
be removed. Note that 0 ≤r≤1, so in the example in Fig. 2, r=2
3= 0.66
(i.e., the element related to MGC value is located in the region R3, so it
does not meet this first requirement);
14 Roberto Saia et al.
2. the value of MGC must be higher than the mean value of the global simi-
larity.
Regarding the first requirement, it should be noted that the regions ex-
tension is strongly related both to the type of items and to their frequency of
fruition, so it depends on the operative scenario. With respect to the second
requirement, we prevent the removal of items when their semantic distance
with the remaining items is lower than the mean value. For this reason, we
first calculate the value of the mean similarity in the context of the user profile,
then we define a threshold value that determines when an item must be con-
sidered incoherent with respect to the current context. Equation 9 measures
the mean similarity, denoted by GS, by calculating the average of the Global
Similarity (GS) values, which are obtained as simW P (sdi, Mu\sdi),∀i∈Iu.
GS =1
|Iu|·X
i∈Iu
(simW P (sdi, Mu\sdi)) (9)
where |Iu|represents the total number of items stored in the profile (in
the case of sample data shown in Table 1, the GS = 0.2906). Once this average
value is obtained, we can proceed to define the condition ρto be used to decide
when an item has to be removed (1) or not (0), based on the average value
GS (as shown in Equation 10).
ρ=(1,if MGC < GS
0,otherwise (10)
Based on the above considerations, we can now define Algorithm 1, used to
remove the semantically incoherent items from a user profile. The algorithm
requires as input the set Iu(i.e., the user profile), and a removal area rused
to define in which part of the profile an item can be removed. In step 3 we
extract the set of synsets Mu(Definition 4) from the description of the items
in the user profile Iu(Definition 2). Steps 4-6 compute the similarity between
each couple of synsets that belong to the user profile. In step 7, the average
of the similarities is computed, so that in steps 8-15 we can evaluate if an
item has to be removed from a user profile or not. In particular, once an item
iis removed from a profile in step 12, its associated similarity sis removed
from the list S(step 13), so that M GC in step 9 can be set as the minimum
similarity value after the item removal. In step 16, the algorithm returns the
user profile Iuwithout the removed items.
4.2.4 Item Recommendation
After the user profile has been processed by Algorithm 1, this step computes
the similarity with all the items not evaluated, and recommends to a user a
subset of those with the highest similarity. An interesting aspect to consider
is that, thanks to our proposal, a user profile in which both the preferences
A Semantic Approach to Remove Incoherent Items From a User Profile 15
Algorithm 1 DCBM Algorithm
Require: Iu=set of items in the user profile, r=removal area
1: procedure Process(Y)
2: N=|Iu|
3: Mu=GetSynsets(Iu)
4: for each Pair p=(sdi, Mu\sdi) in Iudo
5: S←simW P (p)
6: end for
7: a=Average(S)
8: for each sin Sdo
9: MGC =M in(S)
10: i=index(MGC )
11: if i < r ∗size(Iu) AND M GC < a then
12: Remove(i)
13: Remove(s)
14: end if
15: end for
16: Return Iu
17: end procedure
in terms of ratings and in terms of synsets would be available to a recom-
mender system. This would make possible the generation of recommendations
with both collaborative filtering and content-based approaches. In this study
we consider the two main subclasses of collaborative filtering approaches (i.e.,
the neighborhood approach and the latent factor models), since collaborative
filtering approaches are known to be the most accurate. In particular, we con-
sider SVD [25] for the latent factor model-based approaches and a user-based
approach [35] for the neighborhood subclass. This will allow us to evaluate
our approach both in a scenario in which the feature space is reduced (SVD),
and in which the recommendations are instead produced by considering all the
items a user evaluated (user-based approach).
4.2.5 Summary
We can summarize the implementation process of the DCBM algorithm on a
recommender system in the following four steps:
1. Data Preprocessing: preprocessing of the textual description of all items,
in order to remove the useless elements and the items with a rating lower
than the average;
2. Semantic Similarity: WordNet features are used to retrieve, from the
preprocessed text, all the possible pairs between the WordNet synsets in
the text of the items not evaluated and the synsets in the text of the user
profile, keeping only the pairs that have at least an element with the same
part-of-speech, for which we measure the semantic similarity according to
the Wu and Palmer metric;
3. Dynamic Coherence-Based Modeling: the items dissimilar from the
average preferences of a user are identified by measuring the Minimum
Global Coherence (MGC). Moreover, in accordance with certain criteria,
the items that are more semantically distant from the context of a user’s
real preferences are removed from the user profile;
16 Roberto Saia et al.
4. Item Recommendation: the user profile without the incoherent items is
employed to filter the items a user has not yet evaluated, in order to select
the items to recommend.
5 Experimental Framework
This section presents the framework used to evaluate our proposal. The strat-
egy that drove our experiments is first described; then the dataset used and
the preprocessing made on the data are introduced; after, the metrics for the
evaluation are presented; the last part of the section presents the obtained
results.
5.1 Experimental Strategy
The experimental environment for this work is based on the Java language,
with the support of Java API implementation for WordNet Searching (JAWS)
previously mentioned. In order to perform the evaluation, we used two real-
world datasets widely-employed in the recommender systems literature, i.e.,
Yahoo! Webscope (R4) and Movielens 10M.
The recommender systems used to test the effectiveness of the DCBM
algorithms are SVD and a classic User-Based Nearest Neighbors Collaborative
Filtering approach. As previously highlighted, this will allow us to evaluate our
proposal on recommender systems in which the dimensionality is reduced and
in which the feature space is processed as it is.
The Mahout framework was used to implement these recommender sys-
tems. In addition to the training set, the framework requires two parameters
for SVD: the number of target features and the number of training steps to
run (parameter λ). The value of the first parameter can be chosen arbitrarily,
on the basis of the number of features that the SVD should target; in our
experiments the performance differences measured between our approach and
SVD are almost the same by varying this value3; we have set it to 19 for the
Yahoo dataset and to 18 for the Movielens dataset (i.e., as the number of gen-
res of each dataset); the second parameter, λ, will instead be tested through
a set of experiments.
For the user-based approach, it is necessary to set the number of neigh-
bors to consider when predicting the ratings; also this parameter will be set
thanks to a specific set of experiments. The distance function chosen for the
neighborhood approach is the Pearson correlation, one of the most common
measure of correlation to evaluate the similarity between two users [19].
In order to validate our proposal, a comparative analysis has been per-
formed, by considering the values obtained by a recommender system both in
a scenario in which the user profile is processed with the DCBM algorithm
3The analysis has been omitted, since it did not show significant and interesting results,
and in order to facilitate the reading of the paper.
A Semantic Approach to Remove Incoherent Items From a User Profile 17
and in a scenario in which the profile is processed as it is. The comparisons
have been made by measuring both the Root Mean Squared Error (RMSE)
and the Average Difference (AD).
RMSE was chosen as a metric to compare the algorithms because, as the
organizers of the Netflix prize highlight4, it is well-known and widely used,
it allows to evaluate a system through a single number, and it emphasizes
the presence of large errors (both false positives and false negatives). AD was
measured since it is being recently considered by largely employed frameworks
(such as Apache’s Mahout) to facilitate the interpretability of the results [36].
Three sets of experiments have been performed:
1. Analysis of the removed items. We analyze the amount of items re-
moved for each user, in order to analyze the impact of the DCBM algorithm
on a user profile.
2. Recommendation accuracy measurement. We validate the capability
of our approach to remove incoherent items, by analyzing the accuracy of
the generated recommendations with and without the incoherent items in
the user profiles.
3. Performance analysis. We consider the amount of time it takes to pro-
cess an item, when analyzing if it should be removed or not. This allows us
to evaluate our algorithm considering its performance, and to analyze, once
an item gets removed, how much time can be saved by using our proposal.
5.2 Datasets
In order to evaluate our strategy, we perform a series of experiments on two
different real-world datasets, extracted by two standard benchmarks for rec-
ommender systems: Yahoo! Webscope R45and Movielens 10M6.
Yahoo! Webscope (R4). This dataset contains a large amount of data
related to user preferences expressed on the Yahoo! Movies community that
are rated on the base of two different scales, from 1 to 13 and from 1 to 5
(we use the latter). The training data is composed by 7,642 users (|U|),
11,915 movies/items (|I|), and 211,231 ratings (|P|). All the users in the
training set have rated at least 10 items and all items are rated by at least
one user. The test data is composed by 2,309 users, 2,380 items, and 10,136
ratings. There are no test users/items that do not also appear in the training
data. All the users in the test set have rated at least one item and all items
have been rated by at least one user. The items are classified in 19 different
classes (genres), and it should be noted that an item may be classified with
multiple classes. The information of this dataset, training and test data, are
sorted in chronological order, and it should be also noted that the test data
4http://www.netflixprize.com/faq
5http://webscope.sandbox.yahoo.com
6http://grouplens.org/datasets/movielens/
18 Roberto Saia et al.
were gathered chronologically after the training data.
Movielens 10M. The second dataset used in this work is composed by
71,567 users (|U|), 10,681 movies/items (|I|), and 10,000,054 ratings
(|P|). It was extracted at random from MovieLens (a movie recommendation
website). All the users in the dataset had rated at least 20 movies, and each
user is represented by a unique ID. The ratings of the items are based on a
5-star scale, with half-star increments. In this dataset the items are classified
in 18 different classes (movie genres), and also in this case each item may be
classified with multiple classes (genres). Since the Movielens 10M dataset does
not contain any textual description of the items, to obtain this information we
used a file provided by the Webscope (R4) dataset, which contains a mapping
from the movie IDs used in the dataset to the corresponding movie IDs and
titles used in the MovieLens dataset. In this dataset the users were selected
at random for the inclusion in the training and test data.
5.3 Metrics
The accuracy of the predicted ratings was measured through the Root Mean
Squared Error (RMSE) and the Average Difference (AD). Both metrics con-
sider the test set and the predicted ratings, by comparing each rating rui,
given by a user ufor an item iand available in the test set, with the rating pui
predicted by a recommender system. The formulas are shown below, where n
is the number of ratings available in the test set:
RM SE =v
u
u
t
n
P
i=0
(rui −pui)2
n
AD =
n
P
i=0
|rui −pui |
n
5.4 Experimental Results
This section presents the results for each set of experiments presented in the
strategy.
5.4.1 Analysis of the Removed Items
Fig. 3 reports the number of removed items and the amount of users for which
that amount of items were removed, in the Webscope dataset (Fig. 3a) and in
the Movielens dataset (Fig. 3b). Both figures show a power law distribution,
which indicates that for the vast majority of the users just a small number of
A Semantic Approach to Remove Incoherent Items From a User Profile 19
(a) Webscope Profiles Filtering
0 200 400 600 800
0
2,000
4,000
Removed Items
Users Involved
(b) Movielens Profiles Filtering
0 1,000 2,000 3,000
0
2,000
4,000
Removed Items
Users Involved
Fig. 3: Removed Items
items gets removed. However, there are cases in which hundreds of items are
removed from a single user profile and get classified by our approach as noise.
In the next set of experiments we evaluate the capability of our approach
to generate accurate recommendations by employing these profiles without
incoherent items.
5.4.2 Recommendation Accuracy Measurement
Fig. 4 shows the results obtained for the SVD recommender system and the two
employed metrics, in both the Webscope (Fig. 4a-b) and Movielens (Fig. 4c-
d) datasets. The results of the Webscope dataset show that the best results
can be obtained with λ= 0.141, RM S E = 1.0124, AD = 0.7788 for the
unfiltered dataset and λ= 0.171, RM SE = 1.0128, AD = 0.7301 for the data
filtered with the DCBM algorithm. Regarding the results with the Movielens
data, the highest accuracy is obtained with λ= 0.141, RM S E = 0.96, AD =
0.7788 for the unfiltered dataset and λ= 0.091, RM SE = 0.9751, AD =
0.755 for the data filtered with the DCBM algorithm. We also performed
independent-samples two-tailed Student’s t-tests, which highlighted that there
is no statistical difference between the results (p > 0.05). Indeed, we can
notice that the RMSE of the filtered data is slightly higher, while the AD
is slightly lower. This means that, when the feature space is reduced by the
recommendation approach, we can reduce the complexity of the system by
processing less items while keeping the same accuracy.
In Fig. 5 we analyzed the other scenario, in which the feature space pro-
cessed by the recommender system is the entire user profile, by considering
the nearest-neighbors user-based filtering. Fig. 5 shows the results with the
employed metrics, in both the Webscope (Fig. 5a-b) and Movielens (Fig. 5c-d)
datasets. The results of the Webscope dataset show that the best results can
be obtained with neighbors = 8, RM SE = 1.058, AD = 0.7408 for the unfil-
20 Roberto Saia et al.
(a) Webscope Average Difference
0 0.5 1
1
1.5
2
λ
Average Difference
Unfiltered
Filtered
(b) Webscope Root Mean Square
0 0.5 1
1
2
3
4
5
λ
RMSE
Unfiltered
Filtered
(c) Movielens Average Difference
0 0.5 1
1
2
3
λ
Average Difference
Unfiltered
Filtered
(d) Movielens Root Mean Square
0 0.5 1
1
2
3
λ
RMSE
Unfiltered
Filtered
Fig. 4: SVD recommendation accuracy
tered dataset and neighbors = 7, RM SE = 0.8785, AD = 0.5877 for the data
filtered with the DCBM algorithm. Regarding the results with the Movielens
data, the highest accuracy is obtained with neighbors = 98, RM SE = 1.0485,
AD = 0.8255 for the unfiltered dataset and neighbors = 3, RM S E = 1.0035,
AD = 0.7421 for the data filtered with the DCBM algorithm. These results
show that when the entire feature space is considered (as the collaborative
filtering neighborhood-based approach tested in this set of experiments, or the
content-based systems), the recommendation accuracy can be significantly im-
proved (p < 0.05) by removing the incoherent items from the user profiles.
5.4.3 Performance Analysis
Fig. 6 shows the average amount of time in seconds it takes for the DCBM
algorithm to process an item and to decide if it should be removed or not, by
A Semantic Approach to Remove Incoherent Items From a User Profile 21
(a) Webscope Average Difference
0 50 100
0.6
0.7
0.8
0.9
Neighborhood Size
Average Difference
Unfiltered
Filtered
(b) Webscope Root Mean Square
0 50 100
1
1.2
Neighborhood Size
RMSE
Unfiltered
Filtered
(c) Movielens Average Difference
0 50 100
1
1.05
1.1
1.15
Neighborhood Size
Average Difference
Unfiltered
Filtered
(d) Movielens Root Mean Square
0 50 100
1
1.05
1.1
1.15
Neighborhood Size
RMSE
Unfiltered
Filtered
Fig. 5: User-based Approach Recommendation Accuracy
considering the first 1000 user profiles, in order to achieve a statistic relevance.
The dashed line in each subfigure indicates the average time considering all the
user profiles. In the Webscope dataset (Fig. 6a) the average time to process an
item is 0.21 seconds, while for the Movielens dataset (Fig. 6b) it is 0.36 seconds.
This means that the algorithm presents a very good performance considering
the semantic process behind the DCBM algorithm. Moreover, this quick pro-
cess to decide if an item should be removed or not leads to great improvements
when the recommendations are processed, both in terms of number of items to
process and in terms of accuracy when the whole feature space is considered.
22 Roberto Saia et al.
(a) Webscope
0 500 1,000
0
0.2
0.4
0.6
0.8
User Profile
Time per Item (sec)
(b) Movielens
0 500 1,000
0
1
2
User Profile
Time per Item (sec)
Fig. 6: Computation Time
5.4.4 Discussion
We verified that the proposed approach is actually able to remove the in-
coherent items from the user profiles, as highlighted in Fig. 3, which shows
that there are many cases in which tens of items are removed from a single
user profile. Considering that the removed items are not useful to model the
real preferences of the users (i.e., they have been classified as incoherent by
our approach), we obtain a considerable reduction of the computational load,
without any side effect.
The main considerations that arise from the observation of the experimen-
tal results are the following: when we adopt a recommendation approach at the
state of the art that does not modify the feature space, the accuracy improves
significantly, as it happens with the user-based approach, whose results are
shown in Fig. 5; otherwise, when we use an approach that reduces this space,
as happens by using SVD, the accuracy remains unaltered, as we can observe
in Fig. 4.
In this last case, the advantages that come from the use of our approach
are to be found in the reduced number of items taken into account in the
recommendation process. In fact, considering that the removal of the items
from the user profiles does not require much time, as shown in Fig. 6, we have
a considerable reduction of the time needed to perform the recommendation
process, because it operates by analyzing a number of items that is fewer than
the initial one, as reported in Fig. 3.
If on the one hand the proposed approach conducts toward more accurate
recommendations, on the other hand it reduces the number of items in the user
profiles, thus the computational complexity. About this aspect, it should be
observed that the computational load related to the comparisons performed to
detect the MGC do not involve the recommendation process, since it happens
in another moment (i.e., when a user evaluates a new item). Even if our ap-
A Semantic Approach to Remove Incoherent Items From a User Profile 23
proach has a high computational complexity, it is meant to run in background.
Indeed, when a new item is evaluated by the user, it would not be removed
(even if incoherent) since it is recent, so its similarity with the rest of the user
profile can be processed in background. Therefore, the recommendation task,
which is the one that has to run fast because it operates online, can benefit
from the proposed DCBM algorithm and from the removal of the incoherent
items.
In conclusion, our approach improves the performance of a recommender
system, both in terms of accuracy and of reduction of the computational time.
As shown in the results of the experiments, type, and entity of the improve-
ments is related with the recommendation strategy.
6 Conclusions and Future Work
In this paper we proposed a novel approach able to improve the quality of the
user profiling, which takes into account the items related to a user, with the
aim of removing those that do not reflect her real preferences. This is useful in
many contexts, such as when the system does not allow the users to express
their preferences, or when the users do not to make use of this option.
Through it we achieved a twofold result: firstly, we moved the evaluation
process of the items coherence from a domain based on strict mathematical
criteria (i.e., variance of the user’s ratings in the feature space) to a more
flexible semantic domain. Considering that the removal of all incoherent items
from the user profiles leads us toward a considerable reduction of the magic
barrier problem, the second important result is given by the fact that we can
consider each measured improvement as real, instead than a mere overfitting
side effect. The experimental results show that our approach is able to reshape
the user profiles in a coherent way.
A further possible extension might involve the use of a large amounts of
data also related to contexts different from each other as, for example, a sales
platform that gives access to very heterogeneous goods, in which we could
operate in order to discover and process the semantic interconnections between
different classes of items and methods, to evaluate their semantic coherence
during the user profiling activity.
References
1. Addis, A., Armano, G., Giuliani, A., Vargiu, E.: A recommender system based on a
generic contextual advertising approach. In: Proceedings of the 15th IEEE Symposium
on Computers and Communications, ISCC 2010, Riccione, Italy, June 22-25, 2010, pp.
859–861. IEEE (2010)
2. Addis, A., Armano, G., Vargiu, E.: Assessing progressive filtering to perform hierar-
chical text categorization in presence of input imbalance. In: A.L.N. Fred, J. Filipe
(eds.) KDIR 2010 - Proceedings of the International Conference on Knowledge Dis-
covery and Information Retrieval, Valencia, Spain, October 25-28, 2010, pp. 14–23.
SciTePress (2010)
24 Roberto Saia et al.
3. Amatriain, X., Pujol, J.M., Oliver, N.: I like it... I like it not: Evaluating user ratings
noise in recommender systems. In: G. Houben, G.I. McCalla, F. Pianesi, M. Zancanaro
(eds.) User Modeling, Adaptation, and Personalization, 17th International Conference,
UMAP 2009, formerly UM and AH, Trento, Italy, June 22-26, 2009. Proceedings, Lecture
Notes in Computer Science, vol. 5535, pp. 247–258. Springer (2009)
4. Amatriain, X., Pujol, J.M., Tintarev, N., Oliver, N.: Rate it again: increasing recom-
mendation accuracy by user re-rating. In: L.D. Bergman, A. Tuzhilin, R.D. Burke,
A. Felfernig, L. Schmidt-Thieme (eds.) Proceedings of the 2009 ACM Conference on
Recommender Systems, RecSys 2009, New York, NY, USA, October 23-25, 2009, pp.
173–180. ACM (2009)
5. Armano, G., Giuliani, A., Vargiu, E.: 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, pp. 232–237. SciTePress (2011)
6. Armano, G., Giuliani, A., Vargiu, E.: 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, pp. 172–176. IEEE Computer Society (2011)
7. Armano, G., Vargiu, E.: A unifying view of contextual advertising and recommender
systems. In: A.L.N. Fred, J. Filipe (eds.) KDIR 2010 - Proceedings of the Interna-
tional Conference on Knowledge Discovery and Information Retrieval, Valencia, Spain,
October 25-28, 2010, pp. 463–466. SciTePress (2010)
8. Asnicar, F.A., Tasso, C.: ifweb: a prototype of user model-based intelligent agent for
document filtering and navigation in the world wide web. In: Proceedings of Workshop
Adaptive Systems and User Modeling on the World Wide Web’at 6th International
Conference on User Modeling, UM97, Chia Laguna, Sardinia, Italy, pp. 3–11 (1997)
9. Balabanovic, M., Shoham, Y.: Content-based, collaborative recommendation.
Commun. ACM 40(3), 66–72 (1997). DOI 10.1145/245108.245124. URL
http://doi.acm.org/10.1145/245108.245124
10. Bellog´ın, A., Said, A., de Vries, A.P.: The magic barrier of recommender systems - no
magic, just ratings. In: V. Dimitrova, T. Kuflik, D. Chin, F. Ricci, P. Dolog, G. Houben
(eds.) User Modeling, Adaptation, and Personalization - 22nd International Confer-
ence, UMAP 2014, Aalborg, Denmark, July 7-11, 2014. Proceedings, Lecture Notes in
Computer Science, vol. 8538, pp. 25–36. Springer (2014)
11. Billsus, D., Pazzani, M.J.: A hybrid user model for news story classification. In:
Proceedings of the Seventh International Conference on User Modeling, UM ’99,
pp. 99–108. Springer-Verlag New York, Inc., Secaucus, NJ, USA (1999). URL
http://dl.acm.org/citation.cfm?id=317328.317338
12. Budzik, J., Hammond, K.J.: User interactions with everyday applications as context for
just-in-time information access. In: Proceedings of the 5th International Conference on
Intelligent User Interfaces, IUI ’00, pp. 44–51. ACM, New York, NY, USA (2000)
13. Capelle, M., Frasincar, F., Moerland, M., Hogenboom, F.: Semantics-based news recom-
mendation. In: Proceedings of the 2Nd International Conference on Web Intelligence,
Mining and Semantics, WIMS ’12, pp. 27:1–27:9. ACM, New York, NY, USA (2012)
14. Capelle, M., Hogenboom, F., Hogenboom, A., Frasincar, F.: Semantic news recommen-
dation using wordnet and bing similarities. In: Proceedings of the 28th Annual ACM
Symposium on Applied Computing, SAC ’13, pp. 296–302. ACM, New York, NY, USA
(2013)
15. Chirita, P.A., Nejdl, W., Paiu, R., Kohlsch¨utter, C.: Using odp metadata to personalize
search. In: Proceedings of the 28th Annual International ACM SIGIR Conference on
Research and Development in Information Retrieval, SIGIR ’05, pp. 178–185. ACM,
New York, NY, USA (2005)
16. Dennai, A., Benslimane, S.M.: Toward an update of a similarity measurement for a bet-
ter calculation of the semantic distance between ontology concepts. In: The Second In-
ternational Conference on Informatics Engineering & Information Science (ICIEIS2013),
pp. 197–207. The Society of Digital Information and Wireless Communication (2013)
17. Fellbaum, C.: WordNet: An Electronic Lexical Database. Bradford Books (1998)
A Semantic Approach to Remove Incoherent Items From a User Profile 25
18. Finkelstein, L., Gabrilovich, E., Matias, Y., Rivlin, E., Solan, Z., Wolfman, G., Ruppin,
E.: Placing search in context: The concept revisited. ACM Trans. Inf. Syst. 20(1),
116–131 (2002)
19. Herlocker, J.L., Konstan, J.A., Borchers, A., Riedl, J.: An algorithmic framework for
performing collaborative filtering. In: SIGIR, pp. 230–237. ACM (1999)
20. Herlocker, J.L., Konstan, J.A., Terveen, L.G., Riedl, J.: Evaluating collaborative filter-
ing recommender systems. ACM Trans. Inf. Syst. 22(1), 5–53 (2004)
21. Hill, W.C., Stead, L., Rosenstein, M., Furnas, G.W.: Recommending and evaluating
choices in a virtual community of use. In: I.R. Katz, R.L. Mack, L. Marks, M.B.
Rosson, J. Nielsen (eds.) Human Factors in Computing Systems, CHI ’95 Conference
Proceedings, Denver, Colorado, USA, May 7-11, 1995., pp. 194–201. ACM/Addison-
Wesley (1995)
22. Jiang, J.J., Conrath, D.W.: Semantic similarity based on corpus statistics and lexical
taxonomy. arXiv preprint cmp-lg/9709008 (1997)
23. Karypis, G.: Evaluation of item-based top-n recommendation algorithms. In: Proceed-
ings of the 2001 ACM CIKM International Conference on Information and Knowledge
Management, Atlanta, Georgia, USA, November 5-10, 2001, pp. 247–254. ACM (2001).
DOI 10.1145/502585.502627. URL http://doi.acm.org/10.1145/502585.502627
24. Kelly, D., Teevan, J.: Implicit feedback for inferring user preference: a bibliography.
SIGIR Forum 37(2), 18–28 (2003)
25. Koren, Y., Bell, R.M.: Advances in collaborative filtering. In: F. Ricci,
L. Rokach, B. Shapira (eds.) Recommender Systems Handbook, pp.
77–118. Springer (2015). DOI 10.1007/978-1-4899-7637-6 3. URL
http://dx.doi.org/10.1007/978-1-4899-7637-6_3
26. Lam, W., Mukhopadhyay, S., Mostafa, J., Palakal, M.J.: Detection of shifts in user
interests for personalized information filtering. In: SIGIR, pp. 317–325 (1996)
27. Leacock, C., Chodorow, M.: Combining local context and wordnet similarity for word
sense identification. In: C. Fellbaum (ed.) WordNet: An Electronic Lexical Database,
pp. 305–332. MIT Press (1998)
28. Li, L., Yang, Z., Wang, B., Kitsuregawa, M.: Dynamic adaptation strategies for long-
term and short-term user profile to personalize search. In: G. Dong, X. Lin, W. Wang,
Y. Yang, J.X. Yu (eds.) Advances in Data and Web Management, Joint 9th Asia-Pacific
Web Conference, APWeb 2007, and 8th International Conference, on Web-Age Infor-
mation Management, WAIM 2007, Huang Shan, China, June 16-18, 2007, Proceedings,
Lecture Notes in Computer Science, vol. 4505, pp. 228–240. Springer (2007)
29. Lieberman, H.: Letizia: An agent that assists web browsing. In: Proceedings of the
14th International Joint Conference on Artificial Intelligence - Volume 1, IJCAI’95, pp.
924–929. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA (1995). URL
http://dl.acm.org/citation.cfm?id=1625855.1625975
30. Lin, D.: An information-theoretic definition of similarity. In: J.W. Shavlik (ed.) Pro-
ceedings of the Fifteenth International Conference on Machine Learning (ICML 1998),
Madison, Wisconsin, USA, July 24-27, 1998, pp. 296–304. Morgan Kaufmann (1998)
31. Linden, G., Smith, B., York, J.: Industry report: Amazon.com recommendations: Item-
to-item collaborative filtering. IEEE Distributed Systems Online 4(1) (2003). URL
http://dsonline.computer.org/0301/d/wp1lind.htm
32. Liu, F., Yu, C., Meng, W.: Personalized web search by mapping user queries to cate-
gories. In: Proceedings of the Eleventh International Conference on Information and
Knowledge Management, CIKM ’02, pp. 558–565. ACM, New York, NY, USA (2002)
33. Lops, P., de Gemmis, M., Semeraro, G.: Content-based recommender systems: State of
the art and trends. In: F. Ricci, L. Rokach, B. Shapira, P.B. Kantor (eds.) Recommender
Systems Handbook, pp. 73–105. Springer (2011)
34. Ma, Z., Pant, G., Sheng, O.R.L.: Interest-based personalized search. ACM Trans. Inf.
Syst. 25(1) (2007)
35. Ning, X., Desrosiers, C., Karypis, G.: A comprehensive survey of neighborhood-based
recommendation methods. In: F. Ricci, L. Rokach, B. Shapira (eds.) Recommender
Systems Handbook, pp. 37–76. Springer (2015). DOI 10.1007/978-1-4899-7637-6 2.
URL http://dx.doi.org/10.1007/978-1-4899-7637-6_2
36. Owen, S., Anil, R., Dunning, T., Friedman, E.: Mahout in Action. Manning Publications
Co., Greenwich, CT, USA (2011)
26 Roberto Saia et al.
37. Pazzani, M., Muramatsu, J., Billsus, D.: Syskill & webert: Identifying inter-
esting web sites. In: Proceedings of the Thirteenth National Conference on Ar-
tificial Intelligence - Volume 1, AAAI’96, pp. 54–61. AAAI Press (1996). URL
http://dl.acm.org/citation.cfm?id=1892875.1892883
38. Pazzani, M.J., Billsus, D.: Content-based recommendation systems. In: P. Brusilovsky,
A. Kobsa, W. Nejdl (eds.) The Adaptive Web, pp. 325–341. Springer-Verlag, Berlin,
Heidelberg (2007). URL http://dl.acm.org/citation.cfm?id=1768197.1768209
39. Pedersen, T., Patwardhan, S., Michelizzi, J.: Wordnet::similarity: Measuring the re-
latedness of concepts. In: Demonstration Papers at HLT-NAACL 2004, HLT-NAACL–
Demonstrations ’04, pp. 38–41. Association for Computational Linguistics, Stroudsburg,
PA, USA (2004)
40. Pretschner, A., Gauch, S.: Ontology based personalized search. In: 11th IEEE In-
ternational Conference on Tools with Artificial Intelligence, ICTAI ’99, Chicago, Illi-
nois, USA, November 8-10, 1999, pp. 391–398. IEEE Computer Society (1999). DOI
10.1109/TAI.1999.809829. URL http://dx.doi.org/10.1109/TAI.1999.809829
41. Resnik, P.: Using information content to evaluate semantic similarity in a taxonomy.
In: Proceedings of the 14th International Joint Conference on Artificial Intelligence -
Volume 1, IJCAI’95, pp. 448–453. Morgan Kaufmann Publishers Inc., San Francisco,
CA, USA (1995)
42. Ricci, F., Rokach, L., Shapira, B.: Introduction to recommender systems handbook. In:
F. Ricci, L. Rokach, B. Shapira, P.B. Kantor (eds.) Recommender Systems Handbook,
pp. 1–35. Springer (2011)
43. Saia, R., Boratto, L., Carta, S.: Semantic coherence-based user profile modeling in the
recommender systems context. In: Proceedings of the 6th International Conference
on Knowledge Discovery and Information Retrieval, KDIR 2014, Rome, Italy, October
21-24, 2014, pp. 154–161. SciTePress (2014)
44. Said, A., Jain, B.J., Narr, S., Plumbaum, T.: Users and noise: The magic barrier of
recommender systems. In: J. Masthoff, B. Mobasher, M.C. Desmarais, R. Nkambou
(eds.) User Modeling, Adaptation, and Personalization - 20th International Conference,
UMAP 2012, Montreal, Canada, July 16-20, 2012. Proceedings, Lecture Notes in Com-
puter Science, vol. 7379, pp. 237–248. Springer (2012)
45. Said, A., Jain, B.J., Narr, S., Plumbaum, T., Albayrak, S., Scheel, C.: Estimating
the magic barrier of recommender systems: a user study. In: W.R. Hersh, J. Callan,
Y. Maarek, M. Sanderson (eds.) The 35th International ACM SIGIR conference on
research and development in Information Retrieval, SIGIR ’12, Portland, OR, USA,
August 12-16, 2012, pp. 1061–1062. ACM (2012)
46. Salton, G., Wong, A., Yang, C.S.: A vector space model for automatic indexing. Com-
mun. ACM 18(11), 613–620 (1975)
47. Sarwar, B.M., Karypis, G., Konstan, J.A., Riedl, J.: Item-based collaborative filtering
recommendation algorithms. In: V.Y. Shen, N. Saito, M.R. Lyu, M.E. Zurko (eds.)
Proceedings of the Tenth International World Wide Web Conference, WWW 10, Hong
Kong, China, May 1-5, 2001, pp. 285–295. ACM (2001). DOI 10.1145/371920.372071.
URL http://doi.acm.org/10.1145/371920.372071
48. Schafer, J.B., Konstan, J.A., Riedl, J.: Recommender systems in e-commerce. In: Pro-
ceedings of the 1st ACM conference on Electronic commerce, pp. 158–166 (1999)
49. Schickel-Zuber, V., Faltings, B.: Inferring user’s preferences using ontologies. In: Pro-
ceedings, The Twenty-First National Conference on Artificial Intelligence and the Eigh-
teenth Innovative Applications of Artificial Intelligence Conference, July 16-20, 2006,
Boston, Massachusetts, USA, pp. 1413–1418. AAAI Press (2006)
50. Shen, X., Tan, B., Zhai, C.: Implicit user modeling for personalized search. In: O. Herzog,
H.J. Schek, N. Fuhr, A. Chowdhury, W. Teiken (eds.) Proceedings of the 2005 ACM
CIKM International Conference on Information and Knowledge Management, Bremen,
Germany, October 31 - November 5, 2005, pp. 824–831. ACM (2005)
51. Stilo, G., Velardi, P.: Temporal semantics: Time-varying hashtag sense clustering. In:
Knowledge Engineering and Knowledge Management, Lecture Notes in Computer Sci-
ence, vol. 8876, pp. 563–578. Springer International Publishing (2014)
52. Stilo, G., Velardi, P.: Time makes sense: Event discovery in twitter using temporal
similarity. In: Proceedings of the 2014 IEEE/WIC/ACM International Joint Conferences
A Semantic Approach to Remove Incoherent Items From a User Profile 27
on Web Intelligence (WI) and Intelligent Agent Technologies (IAT) - Volume 02, WI-
IAT ’14, pp. 186–193. IEEE Computer Society, Washington, DC, USA (2014). DOI
10.1109/WI-IAT.2014.97
53. Stilo, G., Velardi, P.: Efficient temporal mining of micro-blog texts and its appli-
cation to event discovery. Data Mining and Knowledge Discovery (2015). DOI
10.1007/s10618-015-0412- 3
54. Su, X., Khoshgoftaar, T.M.: A survey of collaborative filtering techniques.
Adv. Artificial Intellegence 2009 (2009). DOI 10.1155/2009/421425. URL
http://dx.doi.org/10.1155/2009/421425
55. Toutanova, K., Klein, D., Manning, C.D., Singer, Y.: Feature-rich part-of-speech tag-
ging 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, pp. 173–180. Association for Computa-
tional Linguistics, Stroudsburg, PA, USA (2003)
56. Vargiu, E., Giuliani, A., Armano, G.: Improving contextual advertising by adopting
collaborative filtering. ACM Trans. Web 7(3), 13:1–13:22 (2013)
57. Wei, C., Khoury, R., Fong, S.: Recommendation systems for web 2.0 marketing. In:
K. Yada (ed.) Data Mining for Service, Studies in Big Data, vol. 3, pp. 171–196. Springer
Berlin Heidelberg (2014)
58. Widyantoro, D.H., Ioerger, T.R., Yen, J.: Learning user interest dynamics with a three-
descriptor representation. JASIST 52(3), 212–225 (2001)
59. Wu, Z., Palmer, M.: Verbs semantics and lexical selection. In: Proceedings of the 32Nd
Annual Meeting on Association for Computational Linguistics, ACL ’94, pp. 133–138.
Association for Computational Linguistics, Stroudsburg, PA, USA (1994)
60. Zeb, M., Fasli, M.: Adaptive user profiling for deviating user interests. In: Computer
Science and Electronic Engineering Conference (CEEC), 2011 3rd, pp. 65–70 (2011).
DOI 10.1109/CEEC.2011.5995827
A preview of this full-text is provided by Springer Nature.
Content available from Journal of Intelligent Information Systems
This content is subject to copyright. Terms and conditions apply.
