Etude de la pertinence de critères de
recherche en recherche d'informations sur
des données structurées
Kris Jack — Florence Duclaye
France Telecom, Research & Development
2 av. Pierre Marzin - 22307 Lannion Cedex - France
kris.jack@orange-ftgroup.com
florence.duclaye@ orange-ftgroup.com
ABSTRACT. An information retrieval system is introduced that takes into account the relative relevance of search
criteria in answer personalisation. The system is evaluated by human participants in the domain of television. When
viewers choose a programme to watch on television, the characteristics of the programme effect their decision. For
example, a person searching for a film may be more interested in the genre than the actors or the director. A study is
conducted that confirms individuals put different relevance on such characteristics. By taking into account the relative
relevance of characteristics, two viewers with exactly the same tastes in features (e.g. instances of actors) but a different
opinion of the relative relevance of characteristics (e.g. actors in general) receive two different search results for the
same query. Accounting for the relevance of search criteria can improve information retrieval.
RÉSUMÉ. Nous présentons un système de recherche d'informations prenant en compte la pertinence relative de critères
de recherche pour la personnalisation de réponses. Ce système est évalué par des participants humains dans le domaine
des émissions télévisées. Lorsqu'un spectateur choisit une émission pour la regarder à la télévision, les caractéristiques
de cette émission influencent sa décision. Par exemple, il se peut que cette personne, qui cherche un film, soit davantage
intéressée par le genre de ce film que par les acteurs ou le réalisateur. Notre étude confirme que les individus attribuent
des pertinences différentes aux caractéristiques. En prenant en compte la pertinence relative de caractéristiques, deux
spectateurs ayant exactement les mêmes goûts en matière de désignations de caractéristiques (par ex. des noms
d'acteurs) mais ayant des opinions différentes sur la pertinence relative des caractéristiques (par ex. les acteurs en
général) reçoivent deux résultats de recherche différents. Le fait de prendre en compte la pertinence relative des critères
de recherche permet d'améliorer les résultats de la recherche d'informations.
KEY WORDS: Information Retrieval; Result Personalisation; Preferences; Characteristic
Relevance; Television.
MOTS-CLÉS: recherche d'informations, personnalisation de résultats, préférences, pertinence
de caractéristiques, télévision.
Title of the journal. Volume X – no X/2002, pages 1 to n

2 Title of the journal. Volume X – no X/2002
1. Introduction
Television viewers want to be entertained. Finding a programme that will satisfy
this need, however, can be difficult. Hundreds of programmes are broadcast over
dozens of channels in a single evening. Traditional searching methods such as
flicking through programming guides (paper-based and electronic) or zapping
through channels can be laborious and time-consuming and far from entertaining.
Furthermore, the viewer may not succeed in finding the most appropriate
programme. A viewer should be able to get from the start state of wanting to be
entertained to the end state of being entertained without the unwelcome searching
caused by information overload in between.
This type of problem has recently attracted much interest within several domains
of information management such as databases (Koutrika, 2005), information
retrieval (Pitkow et al., 2002) and recommender systems (Adomavicius et al., 2005).
Given a request to find an item, that may or may not exist, within a search space, a
system can return a list of items that satisfy the request, to varying degrees. Such
requests are generally more structured in the field of databases than in the other two
fields, with recommender systems often concentrating on general, catch-all requests,
such as “find me something that I would like”. To satisfy users' individual needs,
personalisation has become a popular area of research in all of these fields.
France Télécom Research and Development has developed a semantic
technology that serves as the basis for the system presented in this paper (consult
LePape et al. (2003) for further details). The system will be referred to as FTSEM.
In this paper, FTSEM's functionality is augmented to include result personalisation.
That is, given a request from a user, FTSEM dynamically interprets the request
using data from the user's profile to produce a more personalised result. The style of
result personalisation is based on characteristic relevance. Every characteristic1 that
is used to describe an item has an associated relevance to a given user, with respect
to their needs. For example, in the domain of television, a viewer searching for a
programme may put more relevance on the genre of the programme than on the year
in which it was produced. The relevance of a characteristic thus permeates
throughout all features that belong to that characteristic, that is, all genres are more
relevant than all years of production.
To investigate the personalisation of characteristic relevance, FTSEM is
modified to represent statements such as “When searching for a film, I am more
interested in who directed it than who stars in it”. By taking into account the
relative relevance that people put on characteristics, more personalised search results
can be produced. Consider two people who are searching for a film to watch. Both
of these people have exactly the same preferences in terms of film features. That is,
they both like and dislike the same actors, genres, directors and so on. They differ,
however, in the relative importance that they give to characteristics. One person is
interested in directors above all whereas the other person is interested in genres
1. Terminology: in this paper, characteristic refers to ontological descriptive classes and
feature refers to instance of these classes.

Pertinence de critères en R.I. 3
above all. When searching for a film, a traditional personalised system that is
sensitive to the feature-based preferences would produce the same results for both
people. A system that is sensitive to characteristic relevance, however, is likely to
produce different results for the two users.
In this paper, the notion of characteristic relevance is explored within an
information retrieval system. The applicative domain of the system is television. A
study was conducted that asked whether individuals placed a different amount of
relevance on different characteristics when searching for programmes to watch on
the television. The results of a second study are then reported that considered how
different characteristic relevance ratings could be exploited by FTSEM. Before
formally presenting the system, a number of studies that have informed this research
are discussed.
2. Background
An information system is typically personalised by creating a user profile for
each person who uses the system. When executing a request, the system consults
the user's profile in order to personalise the results. A user profile can be created
and maintained through various methods such as entering data explicitly (Gaush et
al., 2003), machine learning (DeLuca et al., 2005), recording feedback (Yu et al.,
2004) and using dialogue (Krulwich, 1997). The profile itself can record various
facts such as personal data, user preferences, cognitive or learning styles, goal-
related data, level of experience with a system and level of experience in a domain.
By taking into account these different needs, preferences and characteristics, a
system can produce more personalised results.
Preferences have been explored in detail in various domains from philosophy
(Hansson, 2001) and decision theory (Fishburn, 1970) through to artificial
intelligence (Wellman et al., 1991). In the field of databases, preference-based
querying (Keissling, 2002; Chomicki, 2003) allow statements such as “I like A more
than B” to be represented in logic-based formulae. Preferences, however, “are
multiple, heterogeneous, changing (and) even contradictory” (Vallet et al., 2006).
Recent research has demonstrated several useful classifications for preferences
including qualitative or quantitative (Chomicki, 2003), persistent or ephemeral
(Sugiyama et al., 2004), noisy or relevant (Vallet et al., 2006), hard or soft (Berners-
Lee et al., 2001), independent or prioritised (Siberski et al., 2006), present, positive
or negative (Koutrika et al., 2005), present or absent (ibid) and exact or elastic
(ibid). It should be noted, however, that preferences are linked to context and
should be understood in context (Vallet et al., 2006). The relevance of
characteristics, which are the topic of this paper, can be considered as a type of
preference. For example, a user prefers to make his decision based on the genre of a
programme. This preference is closely related to the distinction made between noisy
or relevant preferences by Vallet et al. (2006).

4 Title of the journal. Volume X – no X/2002
Much work in personalisation has been applied to the domains of television and
cinematography. Recommender systems have had much success (see Adomavicius
et al., 2005 for an overview), becoming commercially viable tools. Three main
techniques are employed in these systems; content-based filtering, collaborative
filtering and a hybrid of the two. The work described in this paper is content-based.
Query personalisation in databases has also proved valuable. Koutrika et al. (2005)
show how a user profile can represent statements such as “I am interested in the
director of a film”, a statement similar to the ones investigated in this paper.
3. Personalised Information Retrieval
In this section, a personalised version of FTSEM is formally introduced. It can
represent statements such as “I am more interested in characteristic X than
characteristic Y when I search for an item”. The results of two studies are then
reported and discussed. The first study investigates how people differ in their
preferences towards characteristic relevance and the second study investigates
whether such differences can be exploited in an information retrieval system.
3.1. Personalising FTSEM
FTSEM allows users to retrieve data stored in its database according to semantic
distances and knowledge management algorithms. For the purpose of this paper,
only the database component of FTSEM will be considered. Given a request,
FTSEM determines retrieves a list of ordered items closest to the request. Both the
semantic distance between features and the relevance of characteristics are taken
into account. Semantic distance is dealt with trivially here, as a binary distinction
between equal or not equal. Relevance, on the other hand, is presented in richer
context. The relevance of a characteristic can reinterpret the semantic distance
between features, thus producing a relevant distance.
More formally, FTSEM contains an unordered list of items, I, where I ⊂ D, the
set of all possible data. An item, i, is a vector of features such that i = {f1, f2,… fn}.
The semantic distance between two features f1 and f2 is given by [1].
semantic_distance(f1, f2) = 1 20 if is equal to 1 otherwise
f f [1]
The relevance of a characteristic is expressed as a non-negative integer with an
upper bound of y, {x | x ,∈ ¢ 0 ≤ x < y}. Given the semantic distance between two
features, d = semantic_distance(f1, f2), the relevant distance can be found as a
function of the relevance, μ, of the characteristic by [2].
relevant_distance(μ, d) = upperbound
if is equal to 0 2 1 otherwise
dµ
µ µ

− + [2]

Pertinence de critères en R.I. 5
Thus, if a characteristic is very relevant (e.g. 1 on the scale of 1-10) then the
relevant distance will be low when features are equal and high when features are not
equal. If, on the other hand, the characteristic is not relevant (e.g. 10 on the scale of
1-10) then the relative distance will neither be very low nor very high, regardless of
whether the features are equal or not.
Having defined the relevant distance between features, it is possible to define the
total distance between items. Given two items, i1 = {f1, f2,… fn} and i2 = {fn+1, fn+2,…
fm}, the total distance between these items is given by [3].
total_distance(i1, i2) = j 1+j n+1+j0 j<n
relevant_distance( ,distance( , ))f fµ
≤
∑ [3]
where μj, is the relevance value for characteristic j.
In order to personalise FTSEM, a user profile is stored for each person who uses
the system. The user profile records relevance values for characteristics. The user
profile is thus a list of relevance values. User profiles can be instantiated through
two means. Relevance values can be entered by participants, or they can be learned
by analysing an ordered list of items.
In the first case, users are asked to enter the relevance of each characteristic. For
example, in the domain of television, a set of characteristics may include the
channel, title and length of the programme. For each characteristic, the user is asked
to express how relevant it is to them on a given scale e.g. 1-10. The relevance value
provided is stored for that characteristic in the user's profile.
FTSEM is also able to learn relevance values. To do so, it requires a list of items
ordered by preference, a scale for the relevance values and an indication of which
features are liked and disliked. It is assumed that if liked features appear often and
high in the list, then the characteristic is relevant. Items are ordered so that i1 is the
most liked and item in is the least liked item in the list. The relevance value for each
characteristic is calculated separately. Given characteristic cx, for all features in cx,
the scores of the disliked features are totalled. The score for the disliked feature in
im is then given as n – m + 1 and the relevance for the characteristic is given by [4].
relevance(cx) =
1
total characteristic score scalex
j n
j
≤ ≤
∑ [4]
3.2. The Relative Relevance of Characteristics
It is possible that people differ in the amount of relevance that they place on
different characteristics when selecting a programme to watch on television. For
example, one person may be most interested in the genre of a programme, while
another person may be most interested in the actors. It is not clear, however, if
people really do differ in this respect and how much variety is present. A study was
conducted to explore this issue.

6 Title of the journal. Volume X – no X/2002
3.2.1. Procedure
31 participants were recruited in total. All participants were acquainted with the
problem of searching for a programme to watch. Participants were provided with a
questionnaire. Given 26 characteristics of programmes (e.g. title, channel) they
were asked to mark the relevance (not relevant, relevant or very relevant) of each
characteristic when they search for a programme to watch on television.
Participants marked characteristic relevance with respect to three types of
programmes; films, entertainment and educational and cultural.
3.2.2. Results
All 31 participants completed the task. Each participant rated the relevance of
26 characteristics for all three types of programme, totalling 2418 individual ratings.
Analysis of characteristic ratings shows significant distributional differences
across different programme types. For 15 characteristics, the distributional
difference was not significant across programme types (χ2(4) = 9.49; 0.05). For all
other characteristics, significant distributional differences were found (χ2(4) = 14.86;
0.005). In the majority of these cases (9 out of 11), films were the main contributor.
For one characteristic, cultural and educational programmes contributed the most
and for the remaining characteristic there was no significant single contributor.
Figure 1. Average relevancy scores derived for 26 film characteristics. Thresholds
show the relevance levels between how characteristics are generally perceived.
Each characteristic was given a score based on the relevance ratings that it
received. One point was given to a characteristic each time that it was marked as
relevant, two points when it was marked as very relevant and zero points when it
was marked as not relevant (Figure 1). The general relevance of each characteristic
was also found. For example, if a characteristic was described as not relevant by
more than half of the participants then it was marked as not relevant in general. The
participant's ratings for films were compared individually with the general relevance
Relevance Scores for Film Characteristics
0
10
20
30
40
50
60
Characteristics
very relevant
relevant
not relevant

Pertinence de critères en R.I. 7
of film characteristics. On average, 51% of participants' relevancy ratings were the
same as the general relevance. Similarly, averages of 46% and 49% were found for
entertainment programmes and cultural and educational programmes respectively.
3.2.3. Discussion
The results allow two main conclusions to be drawn. First, there are significant
differences between the perceived relevance of individual characteristics when
people search for different types of programmes. Second, the results confirm that
people differ in the amount of relevance that they place on different characteristics
when selecting a programme to watch on television.
The type of programme that a person wants to watch has an impact on the
perceived relevance of its characteristics. The set of characteristics used, while
applicable to all types of programmes in general, were found to be more relevant for
films. An analysis of relevance distribution reveals that there are significant
differences between the perceived relevance of characteristics across programme
types. Films are the main contributors to this difference. These results can inform
the construction of an appropriate ontology for FTSEM. For example, all
programmes can be described by a genre, but only films have to include a director.
The results also show that a great amount of variety exists across participants'
perceived relevance of characteristics. A general set of characteristic relevance
ratings was generated based on the average scores given by participants. If
participants shared similar perceived relevance ratings then this set would be largely
representative of each participant's ratings. Comparisons, however, find that only
51% of participants' ratings, on average, correspond to the relevance ratings in the
general set. Although this is higher than chance (33%), it highlights that participants
vary in their perception of relevance for characteristics.
3.3. Study to Evaluate Personalised FTSEM
Given that people differ in the relevance that they place on different
characteristics, it is possible to investigate if an automatic information retrieval
system that is sensitive to this preference can exploit it to produce better results. A
study was conducted to explore this potential.
3.3.1. Procedure
24 participants were recruited. All participants were acquainted with the
problem of searching for a programme to watch on television. Participants were
asked to complete a computer-based questionnaire. Each participant was presented
with four film characteristics (genre, channel, director and actors) and was asked to
rank their relevance in terms of not relevant, relevant or very relevant when
choosing to watch them on television. They were asked to rank their relative
relevance from 1-4, where 1 is the most relevant. Participants were then prompted

8 Title of the journal. Volume X – no X/2002
to enter three examples of each characteristic that they liked and three that they
didn't like. These were entered in order of preference.
The questionnaire generated three lists of film descriptions based on the features
entered. In the first list, the film descriptions included every combination of features
that were most liked and most disliked by the participant, producing 16 film
descriptions in total. Similarly, two further lists were generated from the
participant's second and third strongest likes and dislikes. Participants were asked to
order the films in each list in order of their preference to watch them and then state
which ones they would actually like to watch.
FTSEM was tested to determine how well it could predict the ordering of film
descriptions that were produced by participants. A user profile was generated for
each user that contained their likes and dislikes. The lists of film descriptions were
entered, unordered, into FTSEM's database so that the available search data was
equivalent to that provided to participants. FTSEM acquired participant's relevance
ratings for characteristics using two different methods; input of perceived relevance
and learning from list orderings.
3.3.2. Results
22 of the participants completed the study. Two participants failed to complete
the study as they could not name three features that they liked and disliked for each
characteristic. Each participant ranked three lists, producing 66 lists in total.
Participants selected between 1 and 12 of the 16 films to watch, inclusive. In
general, when ordering film descriptions, participants only took care ordering those
films that they would like to have watched. As a result, only the order of films that a
participant would like to watch are considered well ordered. All results presented
are based on these well ordered data.
FTSEM was first tested to determine how well it could predict film description
orderings for specific participants from perceived relevance ratings (Figure 2).
Three different scoring systems were compared; the 3-Scale system (not relevant,
relevant, very relevant), the 4-Scale system (ratings from 1-4) and a combined scale
system (ratings from 1-12). In the combined scale, a score from 1-12 was derived
for the relevance of each characteristic as 4x + y, where x is the score from the 3-
scale system and y is the score from the 4-scale system. In general, the combined
scale scoring system outperforms the other two systems.
FTSEM was tested to determine how well it could predict film description
orderings for the third list, by learning from the first and second lists (Figure 3). A
12-scale scoring system was used for ease of comparison with profiles based on
perceived relevance. The results are based on 10 participants who selected at least
four films to watch in all three of their lists. Three learning strategies are compared;
one which learns from the ordering of list 1, one which learns from the ordering of
list 2 and one which learns from the ordering of both lists 1 and 2.

Pertinence de critères en R.I. 9
Figure 2. Comparison of precision for scoring systems on a 3-point scale, 4-point
scale and 12-point scale. Not all participants chose to watch six films so results
from ranking six films were based on 27 lists, indicated by the number in brackets.
Figure 3. Comparison of precision for learned and perceived relevance. Results
are based on the ten participants who chose to watch at least four films in each list.
Partial profiles and film lists for two participants, who indicated that they would
like to watch the first four films in their lists, are shown (Figure 4). The relevance
values were learned from the participant's first and second lists. Scores for film
descriptions are shown for each participant using both profile A's relevance values
and profile B's relevance values to demonstrate how results differ.
3.3.3. Discussion
The study encouraged participants to both express the perceived relevance of
characteristics and demonstrate their actual relevance in examples. Three different
scoring systems were used that differed in terms of precision for representing
perceived relevance. The best of these scoring systems was used to compare
perceived relevance with learned relevance, using three different sources of training
data. Film descriptions presented to participants were better ordered when using the
Learned Relevance Vs. Perceived Relevance
0%
20%
40%
60%
80%
100%
1 2 3 4
Number of Results
Perceived Relevance
Learned from List 1
Learned from List 2
Learned from Lists 1 & 2
Comparison of Scoring Systems for Perceived Relevance
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1 (66) 2 (52) 3 (41) 4 (40) 5 (33) 6 (27)
Number of Results (Given in Number of Trials)
3-Scale
4-Scale
Combined Scale

10 Title of the journal. Volume X – no X/2002
relevance values from their profile learned from their task-based expressions of
relevance. FTSEM demonstrates that participants are offered films that better match
their preferences when taking into account the values of relative characteristics.
Profile A Actor Channel Director Genre
Relevance Values 5.5 11 0.5 3
Likes Keanu Reeves M6 Steven Spielberg Fantastique
Dislikes Woody Allen France 3 Abel Ferrara Erotique
Six Films Ordered by Participant A
Order Given Actor Channel Director Genre Score A Score B
1 Keanu Reeves M6 Steven Spielberg Fantastique 20 22
2 Keanu Reeves France 3 Steven Spielberg Fantastique 21 23
3 Woody Allen M6 Steven Spielberg Fantastique 32 40
4 Woody Allen France 3 Steven Spielberg Fantastique 33 41
9 Keanu Reeves M6 Steven Spielberg Erotique 37 34
10 Keanu Reeves France 3 Steven Spielberg Erotique 38 35
Profile B Actor Channel Director Genre
Relevance Values 2.5 11 3 5.5
Likes Brad Pitt France 2 Tim Burton Sci-fi
Dislikes Tom Cruise TF1 James Cameron Romance
Six Films Ordered by Participant B
Order Given Actor Channel Director Genre Score A Score B
1 Brad Pitt France 2 Tim Burton Sci-fi 20 22
2 Brad Pitt France 2 Tim Burton Romance 37 34
3 Brad Pitt TF1 Tim Burton Sci-fi 21 23
4 Brad Pitt TF1 Tim Burton Romance 38 35
5 Tom Cruise France 2 Tim Burton Sci-fi 32 40
6 Tom Cruise TF1 Tim Burton Sci-fi 33 41
Figure 4. Two participants' profiles and partially complete lists of ordered film
descriptions. Scores for films are given using profile A's relevance values (Score A)
and profile B's relevance values (Score B). The lower the score, the more the
participant is predicted to like the film. The four lowest scored films are shown in
bold. For participant A, score A gives his order perfectly and score B gives his
first, second, ninth and tenth films. For participant B, score A gives his first, third,
fifth and sixth films and score B gives his first, third, second and fourth films.
Three different scoring systems representing a users' perceived relevance of
characteristics were compared (Figure 2). One system was scored on a 3-point
scale, another on a 4-point scale and the final one on a 12-point scale. The different
systems did not produce significantly different results. In general, however, the 12-
point scale produced the best results. This scale employs the greatest level of detail,
out of the three, suggesting that more detail allows better results to be produced.
FTSEM was tested for how well it could learn a participant's relevance
characteristic preferences from the order in which they ranked lists of films. In
every case (except selection of the first film description, which was equal) relevance

Pertinence de critères en R.I. 11
values from profiles that were learned from users' lists outperformed those that were
instantiated from users' declared preferences. Three sets of relevance values were
learned in different training sessions from user's lists. In general, the relevance
values learned from the ordering of the first list led to the worst of the three
predictions. Relevance values learned from the second list, however, improved
prediction and a relevance values learned from both lists led to the best predictions.
The difference in prediction quality is likely to stem from the varying levels of
extremes present in the data. The first list contains all of a participant's most liked
and most disliked examples of features. To train the system with these extreme
features produces worse results than learning from the less extreme features in the
second list. Finally, a combination of the two lists allows the system to produce a
more balanced representation of the user's actual characteristic preferences.
The results confirm that an automatic information retrieval system that is
sensitive to a person's relative relevance of characteristics can exploit this
information to produce better results (Figure 3). Participant A's profile indicates that
he finds the director of a film most relevant, then the genre, the actors and finally the
channel, whereas participant B's profile indicates that he finds the actors most
relevant, then the director, the genre and the channel. When film ordering
predictions are made using the participants' own profile, their four favourite films
are retrieved. If relevance values from profiles are switched, however, incorrect
predictions are made. The system successfully learns user's tastes with respect to
characteristic relevance.
This study allows insights to be made into the effects of changing characteristic
relevance. When an item's characteristic is very relevant to a decision, a single
feature of that characteristic can clinch the decision. When it is not relevant,
however, the influence of its features is diminished. The exploitation of
characteristic relevance is possible even when a person's feelings towards individual
features are represented in a simply binary, like or dislike, form. In an environment
where the search space is large and diverse, with few features that invoke extreme
preferences, systems can benefit from being sensitive to characteristic relevance.
4. Conclusions
Characteristics have different levels of relevance for different individuals. A
system that is sensitive to this fact can exploit it to personalise its behaviour towards
users. Two users who share the same feature-based preferences but differ in their
attitudes towards characteristic relevance should be differentiated by the system.
Given the same request from these two users, a different set of results should be
produced. The system presented here demonstrates the value of recognising
characteristic relevance in information retrieval.

12 Title of the journal. Volume X – no X/2002
5. References
Adomavicius, G., Tuzhilin, A. “Toward the next generation of recommender systems”, IEEE
Transactions on Knowledge and Data Engineering, vol. 17, no. 6, 2005, p. 734-749.
Berners-Lee, T., Hendler, J., Lassila, O., The Semantic Web, 2001.
Chomicki, J., “Preference formulas in relational queries”, ACM Transactions in Database
Systems, vol. 28, No. 4, 2003, 427-466.
DeLuca, E.W.D., Nümberger, A., “A meta search engine for user adaptive information
retrieval interfaces for desktop and mobile devices”, 10th International Conference on
User Modeling UM'05, Workshop on New Technologies for Personalized Information
Access, 2005, Edinburgh, Scotland, p. 23-33.
Fishburn, P., Utility Theory for Decision Making, New York, Wiley, 1970.
Gaush, S., Chaffee, J., Pretschner, A., “Ontology-based personalized search and browsing”,
Web Intelligence and Agent Systems, vol. 1, issues 3-4, 2003, p.219-234.
Hansson, S.O., “Preference Logic” In Gabbay, D., Guenthner, F. (eds.), Handbook of
Philosophical Logic, vol. 4, chapter 4, 2002, Kluwer, p. 319-393.
Keissling, W., “Foundations of Preferences in Database Systems”, 28th International
Conference on Data Engineering, Hong Kong, China, 2002, p. 311-322.
Koutrika, G. Ioannidis, Y., “Personalized queries under a generalized preference model”, 21st
International Conference on Data Engineering ICDE'05, 2005, Washington, USA, p.
841-852.
Krulwich, B., “Lifestyle finder: Inteligent user profiling Rusing large-scale demographic
data”, AI Magazine, vol. 18, no. 2, 1997, p. 37-45.
LePape, J.-P., Bretier, P., “Technolgie d'exploitation intelligente de connaissances pour la
production de réponses coopératives”, CITE'03, Décembre, 2003.
Pitkow, J., Schutze, H., Cass, T., Cooley, R., Turnball, D., Edmonds, A., Adar, E., Breuel, B.,
“Personalized Search”, Communications of the ACM, vol. 45 , issue 9, 2002, p. 50-55.
Siberski, W., Pan, J.Z., Thaden, U., “Querying the Semantic Web with Preferences”,
International Semantic Web Conference, 2006, p. 612-624.
Sugiyama, K., Hatano, K., Yashikawa, M., “Adaptive Web search based on user profile
constructed without any effort from users”, 13th Conference on World Wide Web,
WWW'04, New York, USA, ACM Press, 2004, p. 675-684.
Vallet, D., Fernandez, M., Castells, P., Mylonas, P., Avrithis, Y., “Personalized Information
Retrieval in Context”, 21st National Conference on Artificial Intelligence, 3rd
International Workshop on Modeling and Retrieval of Context, Boston, 16-17 July, 2006.
Wellman, M.P., Doyle, J., “Preferential semantics for goals”, Ninth National Conference on
Artificial Intelligence AAAI'91, Anaheim, California, USA, 1991, AAAI Press, p. 698-
703.
Yu, Z., Zhou, X., “TV3P: an adaptive assistant for personalized tv”, Consumer Electronics,
IEEE Transations on, vol. 50, vo. 1, 2004, p. 393-399.