=Paper=
{{Paper
|id=Vol-461/paper-4
|storemode=property
|title=Personalized Information Retrieval Approach
|pdfUrl=https://ceur-ws.org/Vol-461/paper4.pdf
|volume=Vol-461
}}
==Personalized Information Retrieval Approach==
https://ceur-ws.org/Vol-461/paper4.pdf
Personalized Information Retrieval Approach
Myriam Hadjouni1, Mohamed Ramzi Haddad1, Hajer Baazaoui1, Marie-Aude
Aufaure2, and Henda Ben Ghezala1
1
Laboratory RIADI-GDL, National School of Computer Sciences, University of Manouba,
2010 la Manouba, Tunisia
{myriam.hadjouni, hajer.baazaouizghal, henda.benghezala}@riadi.rnu.tn
haddad.medramzi@gmail.com
2
MAS Laboratory, SAP BusinessObjects Chair, Ecole Centrale Paris, Grande Voie des
Vignes, 92 295 Chatenay-Malabry, France
marie-aude.aufaure@ecp.fr
Abstract. Web information system personalization is an emergent research
field with the objective to facilitate the use and the control of Web content. This
paper presents a personalized information retrieval approach based on end user
modelling. The proposed approach personalizes data retrieval using implicit
user information and interests measurements. As the data manipulated is
expressed by attributes and values, we define several similarity measures. These
measurements consider both semantic and spatial user contexts. The approach
personalizes Web content and especially spatial information focusing on its
spatial semantic aspects.
Keywords: Web personalization, spatial Web personalization, user modeling
1 Introduction
The volume of information available on Web information systems is growing
continuously. Browsing this content becomes a tedious task given the presentation of
data that does not meet user’s aims and needs. To satisfy user needs, personalization
is an appropriate solution to provide an adaptive and intelligent Human-Computer-
Interaction (H.C.I) and to improve the information systems usability. Moreover, Web
resources increasingly contain geo-referenced entities generally associated with a
geographical location [1]. Geographic information systems suffer from a lack of data
quality since the presented data is highly complex and diverse. Manipulated objects
have a very rich semantic and the degrees of user’s interests towards them vary
depending on context and on personal tastes. Existing personalization approaches try
to determine user preferences to help him while exploring information. However,
these approaches have some drawbacks and generally neglect semantic and spatial
aspect of the manipulated data. The semantic Web provides an appropriate
Proceedings of WISM 2009
background to define and to describe information systems’ content in order to take
into account its different semantics and to efficiently understand it.
This paper presents a personalized information retrieval approach based on end
user modeling. The proposed approach personalizes data retrieval using implicit user
information and interests measurements. We start, in the next section with related
works presentation and discussion. We then present our architecture including user
and data modeling approaches and the similarity measures used to increase the quality
of the personalization process and the measures used to deduce user’s interest. These
measures help us to construct a user’s network based on our proposed system. Section
4 presents this network. Section 5 concludes our work and gives some perspectives.
2 Related Works
Generally, personalization methodologies are divided into two complementary
processes which are (1) the user information collection, used to describe the user
interests and (2) the inference of the gathered data to predict the closest content to the
user expectation. In the first case, user profiles can be used to enrich queries and to
sort results at the user interface level [2]. Or, in other techniques, they are used to
infer relationships like the social-based filtering [3] and the collaborative filtering [4].
For the second process, extraction of information on users’ navigations from system
log files can be used [5]. Some information retrieval techniques are based on user
contextual information extraction [6]. Information semantics are also used to enrich
the personalization process; queries can be enriched by adding new properties from
the available domain ontologies [7]. The user modeling based on ontology can be
coupled with dynamic update of user profile using results of information-filtering and
Web usage mining techniques.
With the evolution of the internet, Web resources increasingly contain geo-
referenced entities generally associated with a geographical location [1]. Statistics
collected through search engines show that spatial information is pervasive on the
Web and that many queries contain spatial specifications, but it is more difficult to
find relevant resources responding to query including a spatial component [8]. The
spatial information personalization should consider spatial properties and
relationships found in Web documents. Design of spatial Web applications requires at
least three components: (1) a user model and associated user preference elicitation
mechanisms and (2) a personalization engine combining spatial and semantic criteria
and (3) a user interface enriched with spatial components [9]. The spatial Web
personalization requires the representation of user features, particularly those relevant
to the spatial domain. [10] explores semantic similarity and spatial proximity
measures as well as relevance ranking functions on the behalf of the user. Semantic
similarity is the evaluation of semantic links existing between two concepts [11]. [12]
introduced a classification algorithm for measuring spatial proximity between two
regions. Another aspect of spatial Web personalization techniques concerns
interactive adaptive map generation and visualization. These techniques are
concerned with Web maps adaptation according to user’s needs [13]. In a prototype
Proceedings of WISM 2009
applied to maritime navigation, [14] categorizes the users according to their
geographical context.
The presented personalization approaches have contributed to the improvement of
information systems use. However and despite their widespread use, these approaches
have weaknesses and limitations. In fact, several approaches, like the collaborative
ones, present the same recommendations for all users within the same cluster. Thus,
they do not consider some specific users preferences when they represent a minority
in a given group. Content based approaches facilitate items retrieval by proposing
some alternatives and recommending similar items to the one that the user is visiting.
However it focuses only on the user’s actual and temporary needs and can’t highlight
the items that are related to the current query results. Other approaches try to
determinate the interests of each user but they are limited by their items model that
doesn’t describe the differences between items properties. This lack of semantic
description of the items decreases the quality of personalization since similarities and
dissimilarities between items can’t be measured accurately. In addition, in most
personalization approaches, the spatial aspect is not taken into consideration, which
requires an adaptation of those approaches to be relevant while applied to spatial
information. These limitations explain the importance given to hybrid approaches.
The hybridization of existing approaches is presented as an alternative that would
improve the quality of personalized systems [15].
3 Personalized Information Retrieval System
We present in this section a Personalized Information Retrieval approach for
Web information systems. Our proposition is based on a dynamic and iterative
construction of a multidimensional user model. This multidimensional approach is
used to represent and describe the user towards different dimensions. The model
proposed (noted Mu) is composed of 4 dimensions: user profile, spatial model, graphic
model and textual model. If we consider U as the set of users, a user u ∈ U will have
as model Mu:
M
u
= D ∪D ∪D ∪P
t s n
(1)
Where:
− Dt = keywords employed by the users for their textual search
− Ds = spatial positions of the users
− Dn = entities visited by the users
− P = user profile.
This model is part of the user modeling level of our system which is composed of
the following ones: user level, application level, user modeling level, information
retrieval level and storage level [16]. In this paper, we only consider the modeling
level which is also concerned with the exploration and the identification of semantic
and spatial relationships between entities extracted from log files or from real-time
navigation.
Proceedings of WISM 2009
Users can be known or unknown. A known user has an account and is individually
identified by the system while an unknown user does not possess proper profile. So,
the user modelling level is also dealing with:
− – For unknown users: (1) the construction of user real-time profile and (2) the
mapping of this current user into the closest prototype.
− – For known users: (1) the activation of the concerned user model and (2) the
update of his real-time profile.
The model is based on an implicit interaction with the user: implicit because the
user isn’t directly asked to give opinion, and interactive because we use the
navigation to measure its interest to a given entity. These measurements are based on:
− The similarities that could exist between attributes and entities of interest.
− The deduction of user interests from all its navigation.
− The calculation of the pertinence of a supposed spatial move. In fact, in our
approach, we also consider that if a user in interested on a spatial zone, he aims
to move to.
Next sections present the measurements cited above.
3.1. Similarity measures
The similarity measures used in our system are attributes values similarity and
entity similarity.
Attributes values similarity These attributes can be simple attributes like numeric
values, Booleans and strings. Otherwise, attributes can be composite such as numeric
intervals and sets, or strings sets. We define for each attribute type a similarity
measure:
− Numeric values: These values are discrete, so it is necessary to have a similarity
function to divide their variation intervals into sets of similar values such as for
hotel prices or rooms number. Each value can be considered to be similar to
neighbourhood’s values. The neighbourhood’s width increases when the value is
high and vice versa. We define neighbourhood of similar values as:
⎧ ⎡ ⎛ α ⎞, α + ⎛ α ⎞ ⎤ ⎫
Sim (α ) = ⎨ β , β ∈ ⎢α − ⎜ ⎟ ⎜ ε ⎟⎥ ⎬ (2)
⎩ ⎣ ⎝ε ⎠ ⎝ ⎠⎦ ⎭
Where ε defines the neighbourhood’s width.
Semantic distance between α and β ∈ [a,b] is:
α − β
D
sem
(α , β ) =
a −b
(3)
And the similarity degree between them is:
D
sim
(α , β ) = 1 − D sem (α , β ) (4)
With these three equations, we will have the similarity as:
⎧
⎪
Sim (α ) = ⎨ β , D (α , β ) ≥ 1 −
(2α ε )⎫⎪
sim b−a
⎬ (5)
⎪⎩ ⎪⎭
Proceedings of WISM 2009
− Numeric intervals: The interval attributes may have common values. These values
are included within the similarity measure. If we consider A and B as two numeric
intervals, the similarity degree between them is:
A∩ B
D
sim
( A, B ) =
A∪ B
(6)
− Sets: as for intervals, groups or lists can have common values. Here, we need to
define the well known intersection and union operators to include our similarity.
Let’s consider Sim(Fi,ai) the set of all values that are similar to the value ai of the
attribute Fi. Intersection and union operators taking into account the similarity
applied to the sets A considered as reference, and B are:
Definition 1
A ∩ s B = {a i ∈ A , ∃b j ∈ B tq b j ∈ Sim (Fi , a i )}
A ∪ s B = A ∪ {bi ∈ B , ¬∃ a j ∈ A tq b j ∈ Sim (Fi , ai )}
So, with A and B as two sets of numeric values, the degree of similarity becomes:
{
card A ∩ B }
{ }
s
D
sim
( A, B ) =
card A ∪ B (7)
s
Thus, A and B are considered similar if D ( A, B ) ≥ ε .
sim
ε is the similarity threshold to be determined in relation to the attribute semantic.
− Other types: On the cases of the other types of attributes like strings or Boolean the
similarity operator is equality.
Entities similarity The similarity measure between two entities corresponds to the
aggregation of their attributes similarity degrees. If x and y are two entities of the
same concept, γFi is the attribute’s importance coefficient. αi and βi are the attribute
values submitted by x and y. Semantic distance between them is:
D
sem
(x, y) =
n
∑ γ F i * D sem α i , β i ( ) (8)
i = 1
With n: number of attributes describing x and y which depends on the concept to which they
belong.
Having εi as the similarity entry of the attribute Fi, we consider that x and y are similar
if:
n
D ( x, y ) ≤ ∑ γ Fi * ε i (9)
sem
i =1
So the set of similar entities to x is:
⎧ n ⎫
Sim( x ) = ⎨ D ( x, y ) ≥ 1 − ∑ γF * ε ⎬ (10)
sim i i
⎩ i =1 ⎭
For building the user model, we start with the profile construction, taking interests
into consideration. This information is deduced by measuring user interests related to
Proceedings of WISM 2009
textual, spatial and navigational dimensions. Next section presents the approach used
to build the user profile using information from navigational dimension.
3.2. User interests
The concept of implicit information gathering adopted is as follow: when a user
touches information, he is implicitly voting for it. Based on this assumption, we
introduce measures that aim to deduce user interests. In the following, u represents a
user; an item visited by u is noted as xi and a concept (accommodation, restaurant,
etc…) to which xi belongs is noted C.
Interests indicators The user profile, in our context, is also used to describe user
interest towards the visited items. We use the following parameters:
− Visits frequency: Represents the number of visits to a given item. This number can
reflect the user interest for an item or a group of items and can be used to
determine the most interesting ones.
− Visits duration: In addition to the visits rate, the visit duration is also significant as
users spend more time reading information perceived as relevant [1].
− Explicit items rating: User’s behaviours do not always reflect his real interest: a
user can visit an item description more than once or spend much time on reading it
just to satisfy curiosity. To avoid such situation, the user can explicitly evaluate his
interest. This evaluation is then used to adjust the previous measures notes.
Visited entities interest The main measures used here to deduce the user interest are
the frequency and the duration of visits. To do this, we use all saved user transactions
and their duration. If the user gives an explicit evaluation on a visited entity, we can
refine these measures. The interest degree Ie of a u towards a visited item x,
considering only the semantic aspect of x, is:
1 ⎛ d (u , x ) n v (u , x ) ⎞⎟
I (u , x ) = ∗ ⎜ v + ∗ V (U , x ) (11)
e 2 ⎜ D (u ) N (u ) ⎟
⎝ v v ⎠
Where
dv(u,x) : Duration of the user visits to x.
Dv(u) : Total duration of the user visits.
nv(u,x) : Number of times the user u visited x.
Nv(u) : Total number of user visits u.
V(u,x) : Average rate assigned by u to x. This note is between 0 (not
interesting) and 1 (Very interesting). The default value is 0.5 if the user
did not rate the item.
Thus, the quality of preferences and interests elicitation process increases when the
user evaluates explicitly visited entities. Taking into account ratings is important
when a user spends a long time visiting an entity that is not interesting for him or
when he visits frequently an item that he is not planning to visit. So, rating this entity
would reduce the influence of the duration and the frequency of those visits.
Proceedings of WISM 2009
Concepts interest Spatial entities are classified into several predefined concepts with
regard to the considered domain. Each spatial concept is defined by several attributes
which are used to describe entities belonging to it. In order to make a personalized
and relevant classification of concepts we measure the user interest towards
spatial/semantic entities belonging to a considered concept. With xi ∈ C, the interest
degree of u towards C is:
I
c
(u , c ) =
n
∑ I
e
(u , x i ) (12)
i = 1
Where
n is the number of items xi visited by u.
Ie(u,xi) : The interest degree of u towards xi.
Features values interest After presenting the metric for classifying concepts by
interests’ degrees, we aim to order entities within the same concept according to their
relevance. This is done with correlation to concepts, i.e. the attribute’s interest degree
value is calculated by considering the interest to entities with similar value.
Consider an attribute Fi ∈ C and vk one possible value of Fi and X(Fi,Sim(Fi,vk)) the set
of all entities visited by the user and which have a similar value to vk. Sim(Fi,vk) is the
set of Fi values that are similar to vk. The similarity operators depend on the attribute
type and semantic. The interest value of u the value vk is:
I ⎛⎜ u , x ⎞⎟
( ∑( )) e ⎝
I (u , F , v ) =
f i k
x ∈ F , Sim F , v
j⎠ (13)
j i i k
The next step is to normalize this interest value to be independent of the elements
number.
I (U , F , v )
f i k
I (U , F , v ) =
v i k
j i
()
∑ v ∈ V F I f (U , Fi , vi ) (14)
Where V(Fi) is the set of all possible attribute Fi values.
Interest prediction As already mentioned, user preferences are introduced by
interests degrees on spatial/semantic concepts and their attributes values are deduced
from user navigations. We use the previous metrics to predict the user’s interest
degree for the non-visited items. So, we assume that:
− The user’s interest degree towards a concept increases with the number of visits to
entities of this concept.
− The more preferred entity values are, the more this entity is likely to be visited.
− The concepts attributes have different importance degrees. So we can distinguish
between the critical attributes and those that do not influence the user choices.
This measure is used to predict the most relevant content the user is searching for.
Moreover, it contributes to refine results of textual search mainly based on the use of
concepts and attributes that will define the domain ontology. We aim to deduce this
ontology from previous (and current) search and navigations and not to explicitly
involve the user. To achieve these points, we present the following formula, which
Proceedings of WISM 2009
predicts the degree of user satisfaction for one entity x depending on its concept C and
its attributes values:
n
S ( x , u ) = I (u , C ) ∗ ∑ α ∗ I (u , F , v ) (15)
c F v i
i =1 i
Where
n is the number of the concept attributes C
v is the value of the attributes Fi presented by the entity x.
Ic(u,C) is the interest degree of u on the concept C.
Iv(U,Fi,v) is the interest degree of u on the value v of Fi.
4 Construction of the user results
Having as objective the amelioration of users’ results, we iterate the construction
of his model. In the end of each iteration, we determine the data pertinence and use
this value in the next iteration. With this hypothesis, two types of information are the
basis of personalization: the location of user at the moment of his search and noted Lu
and the location of his area of interest corresponding to the target search zone noted
Zi. The area of interest is deduced from the position on a map in the user interface or
from the search keywords. The results provided to users are constructed on the basis
of three points: (a) the path needed to reach the area of interest Zi from the user
location when searching Lu, (b) the arrival point and (c) the semantic and spatial
distance between a departure point and the objects of interest.
The area of interest path. The estimation of the arrival path over the area of interest
is a quality criterion. In fact, the taken path has an impact on the arrival point into the
area of interest and therefore affects the location of objects of interest displayed to the
user. In this context, we use a qualitative evaluation based on an explicit weighting of
criteria such as cost or type of way. Then, the user can add its own choice value. The
minimum criteria that we have taken are the type of road, its cost and distance
travelled. The weights are made of values between 0 and 1, values which are given in
the form of notes by the user. The path evaluation equation is:
EvalPath = α .Type * β .Cost * λDis tan ce (16)
With α, β and λ ∈ ]0,1]
The arrival estimation. The move of the user from its search location to its area of
interest can provide us two types of information: (1) the arrival point information
deduced from the path taken when move and (2) a location marker that is the collapse
point of the user, we called it departure point. This point may be the subway station
closest to the object of interest or the user residence. The distance between the arrival
and the departure points is equal to 1 when these two points are joined. The
hypotheses are:
− Existence of an arrival point in Zi.
Proceedings of WISM 2009
− If user moves to Zi, he has a departure point and a path: d(arrival point, departure
point)
− The evaluation is based on the path and on the semantic of the searched object:
Eval ( ArrPt , Path ) = D ( ArrPt , DepPt ) * Sem (SearchObje ct ) (17)
SemSpa
Semantic and spatial distances. These distances are the key features to include the
spatial aspect into the information filtering process. Semantic and spatial distances
between the departure point DepPt and the interest objects objecti are evaluated as
shown below:
n
DsemSpa ( ArrPt , DepPt ) = ∑ Dspa(objec ti, DepPt) * Dsem(objec ti, DepPt) (18)
i =1
Where:
n is the number of the interesting objects in the area Zi.
Dspa(objecti,DepPt) is the spatial distance between objecti and DepPt.
Dsem(objecti,DepPt) is the semantic distance between objecti and DepPt.
In the end of every iteration, these three measures are applied to ameliorate the next
one. Taking into account the area of interest path will explicitly implicate the user,
while the arrival estimation and the semantic and spatial distances calculation is
implicitly done.
5 Conclusion and Future Work
Web personalization attracts rising research efforts to facilitate Web information
retrieval and navigation. Generally, Web personalization and user modeling
approaches do not focus on the spatial aspect of the information and the constraints it
implies. In this paper, we have presented some background knowledge on existing
Web and spatial Web personalization systems.
We have then introduced our proposition for a personalized information retrieval
approach which is mainly based on the end user modeling. The user model is
multidimensional and it is built with iterations that use estimations calculated from
implicit user’s navigation information.
Next step of our work is to add to this personalization process (1) interactions
between the users’ models and (2) as a consequence, the construction of a network of
models. In fact, as we consider similarities and distance between the search concepts
and entities within this process, we assume that adding also similarities measurements
between the users’ models could be helpful in the personalization process.
Proceedings of WISM 2009
References
1. Winter, S., Tomko, M.: Translating the Web Semantics of Georeferences. In Taniar, D.;
Rahayu, W. (Eds.), Web Semantics and Ontology, pp. 297-333, Idea Publishing (2006)
2. Koutrika, G., Ioannidis, Y.: A Unified User-profile Framework for Query Disambiguation
and Personalization. In: Workshop on New Technologies for Personalized Information
Access, held in conjunction with the 10th International Conference on User Modeling, pp.
44–53 (2005)
3. Mladenic D.: Text-learning and Related Intelligent Agents: a Survey. IEEE Intelligent
Systems, 14(4):44–54 (1999)
4. Goldberg, K., Roeder, T., Gupta, D., Perkins, C.: Eigentaste, a Constant Time Collaborative
Filtering Algorithm. Information Retrieval Journal, 4(2):133–151, (2001)
5. Paulakis, S., Lampos, C., Eirinaki, M., Vazirgiannis, M.: Sewep: A Web Mining System
Supporting Semantic Personalization. In: 15th European Conference on Machine Learning
and 8th European Conference on Principles and Practice of Knowledge Discovery in
Databases, LNCS, vol. 3202, pp. 552-554, Springer (2004)
6. Jones, G.J.F., Brown, P.J.: Context-Aware Retrieval for Ubiquitous Computing
Environments. Invited paper in Mobile and Ubiquitous Information Access, LNCS, vol.
2954, pp. 227–243. Springer (2004)
7. Messai, N., Devignes, M.D., Napoli, A., Smaïl-Tabbone, M.: Méthode Sémantique pour la
Classification et l’Interrogation des Sources de Données Génomiques. Revue des Nouvelles
Technologies de l’Information, Extraction des connaissances : Etat et Perspectives, pp.43-47
(2005)
8. Yang, Y., Aufaure, M.A., Claramunt, C.: Towards a DL-Based Semantic User Model for
Web Personalization. In: Third International Conference on Autonomic and Autonomous
Systems, pp. 61-61. IEEE Computer Society (2007)
9. Kuhn, W.: Handling Data Spatially: Spatializing User Interfaces. In: 7th International
Symposium on Spatial Data Handling, Advances in GIS Research II, vol. 2, pp.13B.1-
13B.23, IGU (1996)
10. Yang, Y.: Towards Spatial Web Personalization. PhD thesis, Ecole Nationale Supérieure
d’Arts et Métiers (2006)
11. Resnik, P.: Semantic Similarity in a Taxonomy: an Information-Based Measure and its
Application to Problems of Ambiguity in Natural Language. Journal of Artificial
Intelligence Research, 11:95–130 (1999)
12. Larson, R.R., Frontiera, P.: Spatial Ranking Methods for Geographic Information Retrieval
in Digital Libraries. In: Heery, R., Lyon, L., (eds.) ECDL. LNCS, vol. 3232, pp. 45–56.
Springer (2004)
13. Maceachren, A.M., Kraak, M.J.: Research Challenges in Geovisualization. Cartography and
Geographic. Information Science, 28(1):3–12 (2001)
14. Petit, M., Ray, C., Claramunt, C.: A User Context Approach for Adaptive and Distributed
GIS. In: 10th International Conference on Geographic Information Science. Springer-
Verlag, LNGC, pp.121-133 (2007)
15. Burke, R.: Hybrid Recommender Systems: Survey and Experiments. User Modeling and
User-Adapted Interaction, 12(4):331–370 (2002)
16. Hadjouni, M., Baazaoui, H., Aufaure, M.A., Claramunt, C., Ben Ghezala, H.: Towards a
Personalized Spatial Web Architecture. In: Workshop Semantic Web meets Geopatial
Applications, held in conjunction with the AGILE International Conference on Geographic
Information Science (2008)
17. Kelly, D., Belkin, N.J.: Reading Time, Scrolling and Interaction: Exploring Implicit
Sources of User Preferences for Relevant Feedback. In: 24th Annual International ACM
SIGIR Conference on Research and Development in Information Retrieval, pp. 408–409,
ACM (2001)