Utilizing Virtual Communities for Information Retrieval and User

Modeling

Azza Harbaoui

, Sahbi Sidhom

, Malek Ghenima

and Henda Ben Ghezala

Department of Computing, Laboratory RIADI, National School of Computer Sciences, Manouba, Tunisia

Laboratory KIWI-LORIA, University of Lorraine, Vandoeuvre, France

Keywords:

Personalized Information Retrieval, User Modeling, User Proﬁle, Virtual Communities.

Abstract:

Internet has become the largest library in human history. Having such a large library made the search process

more complicated. In fact, traditional search engines respond users by sending back the same results to

different users having expressed different information needs and different preferences. A signiﬁcant part of

difﬁculties,report to vocabulary problems (polysemy, synonymy...). Such problems trigger a strong need to

personalize the search results based on user preferences. The goal of personalized information is to generate

meaningful results interesting to a number of information users using their proﬁle. This paper presents a

personalized information retrieval approach based on user proﬁle. User proﬁle is built from the acquisition of

explicit and implicit user data. The proposed approach also presents a semantic-based optimization method

for user query. The system uses user proﬁle to construct virtual communities. Moreover, it uses the user’s

navigation data to predict user’s preferences in order to update virtual communities.

1 INTRODUCTION

This paper deals with the personalized information re-

trieval (IR). In fact, classic IR methods usually in-

tended for simple textual search, faced new hetero-

geneous documents and rich scalable contents. Con-

sequently, the user is facing these evolutions and is

being more and more unsatisﬁed, searching for IR

search results quality. In order to overcome the lim-

its of existing IRS, the main goal of personalized IR

is to propose effective solutions, by focusing on the

most relevant results to a user query. In this paper, we

propose a novel approach able to capture what is rel-

evant to a user. The proposal relies upon four compo-

nents: (i) a process of disambiguation based on user’s

votes and domain ontology, (ii) multidimensional pro-

ﬁle modeling based on explicit and implicit user’s col-

laboration, (iii) construction of user’s proﬁle network,

(vi) construction of user’s inﬂuence network based on

user’s rate conﬁdence and user’s rate share.

2 RELATED WORKS

The main interest of personalized systems when ex-

tracting information is the use of a process that con-

siders end-user’s interests and preferences. The per-

sonalization process needs representing, accessing

and storing a users personal information. We present

in this section an overview of personalization and user

proﬁle.

2.1 Personalization

In fact, existing personalization approaches (Zhang

et al., 2014) have contributed to the improvement

of information systems use. A large body of re-

search in Information Retrieval (IR) has highlighted

that relevance is a complex and a challenging concept.

However and despite their widespread, the underlying

complexity stems mainly from the fact that relevance

is estimated by considering multiple dimensions and

that most of them are subjective since they are user-

dependent. To answer this challenge, the commonly

used approaches are based on speciﬁc user’s prefer-

ences.

Positioning ourselves under a contextual or even

personalized IR, access to a resource that is relevant

and adapted to user context is a factor of Personal-

ized Information Retrieval Systems (van Rijsbergen,

2013). Involving users in the search process requires

modeling their proﬁles. Indeed, it was found that

among the reasons behind lack of performance of the

used personalization techniques is the integration of

Harbaoui, A., Sidhom, S., Ghenima, M. and Ghezala, H.

Utilizing Virtual Communities for Information Retrieval and User Modeling.

In Proceedings of the 12th International Conference on Web Information Systems and Technologies (WEBIST 2016) - Volume 2, pages 29-34

ISBN: 978-989-758-186-1

 2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reser ved

user proﬁle regardless of the context of use. How-

ever, users are different in nature, with roles and di-

versiﬁed skills ; their preferences can be general, sta-

ble or recurrent. Any information forming a proﬁle

may not be appropriate for all search circumstances.

Personalization systems often focus on partial aspects

of users, and in this case the system uses explicit ac-

quisition, where users express their personal data or

preferences through the speciﬁcation of their inter-

ests. This has the disadvantage of requiring users to

get involved in the process, which imposes on them

an additional charge on their use of the system. In

addition, users, forced for example to ﬁll out a form

or answer a question, can introduce erroneous infor-

mation, which may lead to insigniﬁcant results. Thus,

the current IR problem faces real challenges that are

closely related among others to the size of the docu-

ments on the web. The ﬁrst challenge is to ﬁnd an ef-

ﬁcient way to provide users with the information they

really need. As a result, several studies have consid-

ered integrating user proﬁle in the search process. For

ten years, research (Maleszka, 2015) on IR has been

evolving and has been largely inﬂuenced by an inter-

est in this area, wary to add a little more intelligence

for improved relevance.

2.2 User proﬁle

Modeling the user is at the center of the implemen-

tation of a personalized information retrieval process.

The goal of user modeling is to select the most rele-

vant information that reﬂects user’s interests (Min and

Jones, 2011). This modeling consists of designating a

structure in which we store information that describes

essentially :

• User interests ;

• Preferences ;

• Context ;

• Expected goal of the search;

• Individual traits;

• Experience.

There are several deﬁnitions of user modeling in

the literature (Cheung et al., 1998), (Tanudjaja and

Mui, 2002). We retain some below:

”A user model is a knowledge source in a natural-

language-dialogue system which contains explicit as-

sumptions on all aspects of the user that may be rele-

vant for the dialogue behavior of the system” (Esparza

et al., 2012)

”User model is an explicit representation of the sys-

tem of a particular user’s characteristics that may

be relevant for personalized interaction.” (Wen et al.,

2004)

”The process of gathering information about the users

of computer systems (Treur and Umair, 2011)and of

making this information available to systems which

exploit it to adapt their behavior or the information

they provide to the speciﬁc requirements of individ-

ual users has been termed as user modeling.”

Several techniques were developed in the litera-

ture to model the user. They differ according to the

approach of proﬁle representation and construction

(Micarelli et al., 2007). As our main objective is

to provide the end user with personalized search re-

sults, the proposed approach ﬁrst builds user proﬁle.

It connects then user proﬁles to each other to con-

struct the users virtual communities. The assump-

tion is that when searching for a relevant document,

the search system should use in addition to speciﬁc

user’s needs and previous searches, knowledge on

other users. Search personalization is achieved by re-

turning to the user ranked results using their proﬁle

and query.

As for our objectives, we propose an approach

(cf. Figure 1) with four main components: (1) a user

proﬁle construction component, (2) an acquisition of

user’s data navigation component, (3) a virtual com-

munities construction component and (4) an inﬂuence

network component.

Figure 1: Proposed approach.

User proﬁle is based on an implicit interaction

with the user: implicit because the user is not directly

asked to give opinion, and interactive through using

navigation to measure interest in a given query.

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

2.3 Virtual Communities based User

Proﬁle

The social aspect of user’s personal life is becoming

increasingly important to examine. We deﬁne a vir-

tual community as a group, often informal, of indi-

viduals with a common interest in sharing knowledge

with members of the same community. They can view

the search history, documents relating to previous vis-

its, tags as well as part of the user proﬁle visible to the

public. A virtual community may be represented in

many ways, we show in what follows its most com-

mon graphic representation (Yakoubi and Kanawati,

2013): a community represented by a graph is a set of

points, some of which are directly connected by one

or more links. It can be represented in many ways:

networks waves ; hypertext Graphs; and social net-

works. In our paper, we used graphs to represent the

user’s network. Indeed, each node is simply a user

proﬁle. Each arc represents a relationship between

two users. Each edge is weighted by a weight rep-

resenting the degree of similarity between the edges

previously connected by this arc. We also used com-

munity detection to identify the community to which

a user of our system belongs. Automatic identiﬁca-

tion of communities has attracted much attention in

recent years and many algorithms have been proposed

to model them. These algorithms consider only net-

works structural data regardless of any other attributes

speciﬁc to edges for example. Most communities de-

tection methods are based on the intuition that the

community structure of ﬁeld graphs is naturally a hi-

erarchical structure (a community consists itself of

sub-communities that consist themselves of sub-sub

communities and so on). In addition, if all the meth-

ods do not explicitly target maximization of modu-

larity it is mainly used to compare the performance

of methods. We opted for the approach of ”Louvain”

(Newman, 2004)to detect communities.

2.3.1 Louvain’s Method

Louvain’s method (Newman, 2004) has signiﬁcant

advantages compared to other community detection

methods. It has been validated and tested success-

fully on several projects, including its speed (although

this is not proven, the algorithm has a linear function

or almost), allowing it to process graphs up with bil-

lions of links, its multi-scale aspect, which enables

it to discover communities at different scales, and its

excellent accuracy compared to other methods. For

these reasons, we used this method to detect virtual

communities.

2.3.2 Construction of User’s Inﬂuence Network

An inﬂuence network is a virtual community consist-

ing of user proﬁles, some of which may be inﬂuen-

tial proﬁles. These members, through their proﬁle, al-

low for inﬂuencing members who belong to the same

community, and thus inﬂuence their research. Ac-

cordingly, the virtual community is reduced to a sub-

community or even an inﬂuence network where the

search space is reduced, thereby improving the search

results. To build this network, we deﬁned two vari-

ables that we called trust rate and sharing rate. These

will be described in the following section, applying

an iterative two-stage succession.

2.3.3 Sharing Rate

Sharing rate, denoted ω

, is a numeric value that

ranges between 0 and 1. speciﬁes sharing rate of a

document by any user within his/her community. ω

determines thus the rate of occurrence of a document

shared by a given user in the community to which

he/she belongs. Its formula is deﬁned as follows:

number f riends user(i)

total number user

s community

(1)

2.3.4 Trust Rate

Under certain conditions, a community is the focus of

a user. Trust level of a user j in a user i is denoted by

ci j

and given by the following formula:

ci, j

∑

j=0, i6= j

Ω

i, j

n − 1

(2)

2.3.5 Inﬂuence Network Constructing Steps

The construction of an inﬂuence network allows for

extracting in a clear way the various friendship links

between a user as represented by his/her proﬁle and

other members of the community. To do this, we used

the following approach: For a given user U, we ex-

tract the network of all his/her friends as a ﬁrst stage

and then, for each proﬁle of the community, we de-

ﬁne an inﬂuence rate, deduced from equations. The

system derives the nature of the relationship that links

user U to his/her friends. This deduction is based on

thresholds according to well-deﬁned criteria, namely;

• If inﬂuence rate > 0,9 : this friend is dominant.

• If 0.6 <inﬂuence rate 0,9 : this is a helping Friend.

• If 0.3 <inﬂuence rate <0,6 : this is a cooperative

friend.

Utilizing Virtual Communities for Information Retrieval and User Modeling

• If inﬂuence rate <0,3 : this is a submissive friend.

We report in the following, an illustrative example

of a user. The output of the system after closure of

friends and the inﬂuence network deduced is shown

in (cf. Figure 2).

Figure 2: Example of Sandydog’s inﬂuence network.

2.4 Case Study

After constructing the inﬂuence network, the user

focuses on his/her dominant friend and launches

a research through a Case-based reasoning process

(CBR). To do this, the user will enter the identiﬁer of

his/her friend, the identiﬁer of his/her query and the

assigned ratings. Thus, the system will return 20 doc-

uments deemed relevant to the dominant friend of the

user. This process is supported by the Jcolibri frame-

work. We opted this object-oriented framework, be-

cause it, inter alia, facilitates the construction of CBR

systems. This framework ensures the management of

the case base (add, edit, and delete cases). Just give

the case a structure (in this case, information about

the dominant friend), and then call these operations

to manage the case base. It also offers a rich set of

predeﬁned, reusable similarity functions to calculate

similarity between ontological concepts.

3 EXPERIMENTATION AND

RESULTS INTERPRETATION

For experimentation and evaluation purposes, we em-

pirically tested our approach on the Book Social

Search Dataset (Koolen et al., 2012). The Dataset

consists of 2,8 million book records from Amazon,

extended with social metadata. First, the experimen-

tal users were asked to notify their corresponding rel-

evant results and then precision, recall, Mean Aver-

age Precision (MAP) were calculated. The evaluation

protocol was designed to tune the experimentation pa-

rameters and then to evaluate the effectiveness of our

personalized approach. It is based on three stages :

The ﬁrst is to measure the impact of a user’s im-

plicit interest in a set of documents. Thus, in addition

to the users explicit knowledge, we were able to high-

light a set of implicit knowledge. The second, in addi-

tion to the query processing process, will measure the

impact of the integration of user proﬁle, on improving

search performance. The last is to test the contribu-

tion of integrating virtual communities on improving

the research process. The objective of these exper-

iments is twofold: ﬁrst, to prove the applicability of

the approach and second to compare and validate each

contribution against the baseline.

3.1 The Scenario Process

We proceed to the evaluation of the personalized in-

formation retrieval approach through the following

stages :

1. Analysis of the query (reformulation and disam-

biguation) by using the rates and tags added by

users after a ﬁrst search. Disambiguation is done

by querying the data dictionary WordNet ;

2. Construction of the user proﬁle by adopting a

multidimensional representation and hybrid ac-

quisition of navigation traces (implicit and ex-

plicit aquisition). Detection of new interests of

users is done by querying the ODP-domaine on-

tology ;

3. Construction of virtual communities through the

acquired user proﬁles and restarting the search

process to assess the impact of integration in the

research process ;

4. Detection of inﬂuence networks (reduced virtual

communities).

This network is integrated back into a process of in-

dexing and searching in order to evaluate the perfor-

mance of our method and its impact on the perfor-

mance of indexing in information search. To vali-

date the proposed approach, we conducted tests, us-

ing the experimental toolkit LEMUR

and the weight-

ing scheme t f ×id f .This conﬁguration is used as the

baseline for our comparative evaluation. Next, we

calculated improvement across the different scenarios

and the baseline system.

3.2 Experimental Results

Through our experiments, we try to evaluate the ef-

fectiveness of our proposal over various real user’s

queries. To estimate the quality of the results ob-

tained under different scenarios, we used standard

measures of precision @X, speciﬁcallyP@10, P@20

andP@30 documents which are respectively average

http://www.lemurproject.org/

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

precision of the 10, 20 and 30 ﬁrst documents re-

turned, the NDCG (Normalized Cumulative Gain)

and MAP (Mean Average Precision) of all 40 selected

queries. For each query, the ﬁrst 100 records are re-

turned by the experimental SRI Lemur and average

precisions are calculated to measure the relevance of

the system. The presented results compare precision,

MAP and NDCG values between the personalized ap-

proach and the baseline. We notice that a precision at

5 of a personalized search is better than the baseline.

The sheet at (cf. Figure 3). presents the search re-

sults quality measurement with virtual communities.

We calculate P@5, P@10, and MAP(cf. Figure 4)

for a set of 40 queries. We can see in this sheet that

P@10 presents more accurate results. Moreover, we

see that performance improvement is better for P@10

or P@15 than P@30 and P@100. This improvement

can be explained by the fact that there is less irrele-

vance when we consider the ﬁrst 10 or 15 results. We

can also conclude that the personalized search includ-

ing virtual communities improves system precision.

Figure 3: Recall/Precision Curve.

Figure 4: MAP histogram.

4 CONCLUSIONS

We began this paper with an overview of personaliza-

tion. Then we presented our proposal of using vir-

tual communities to model users for information re-

trieval. The proposed approach includes the creation

of user proﬁles, the construction of virtual commu-

nities and inﬂuence networks. We conducted an ex-

perimentation and evaluation phase on INEX Book

Social Search Dataset (Koolen et al., 2012). Evalu-

ation shows that the system improves search results

when integrating virtual communities and inﬂuence

networks. However, these results show some limita-

tions. In fact, the process of user proﬁle construction

and that of virtual communities construction are not

synchronized. The two processes cannot take place

simultaneously. Proﬁle construction is always done

in batch mode in the case of a cold start.

REFERENCES

Cheung, D. W., Kao, B., and Lee, J. (1998). Discov-

ering user access patterns on the world wide web.

Knowledge-Based Systems, 10(7):463–470.

Esparza, S. G., OMahony, M. P., and Smyth, B. (2012).

Mining the real-time web: a novel approach to product

recommendation. Knowledge-Based Systems, 29:3–

11.

Koolen, M., Kazai, G., Kamps, J., Doucet, A., and Lan-

doni, M. (2012). Overview of the INEX 2011 Books

and Social Search Track. In Geva, S., Kamps, J., and

Schenkel, R., editors, Focused Retrieval of Content

and Structure : 10th International Workshop of the

Initiative for the Evaluation of XML Retrieval, INEX

2011, volume 7424 of Lecture Notes in Computer Sci-

ence, pages 1–29. Springer.

Maleszka, B. (2015). An adaptation method for hierarchi-

cal user proﬁle in personalized document retrieval sys-

tems. In Intelligent Information and Database Sys-

tems, pages 107–116. Springer.

Micarelli, A., Gasparetti, F., Sciarrone, F., and Gauch, S.

(2007). Personalized search on the world wide web.

In The adaptive web, pages 195–230. Springer.

Min, J. and Jones, G. J. (2011). Building user interest pro-

ﬁles from wikipedia clusters.

Newman, M. (2004). Detecting community structure in net-

works. European Physical Journal, 38:321–330.

Tanudjaja, F. and Mui, L. (2002). Persona: A contextu-

alized and personalized web search. In System Sci-

ences, 2002. HICSS. Proceedings of the 35th Annual

Hawaii International Conference on, pages 1232–

1240. IEEE.

Treur, J. and Umair, M. (2011). An agent model integrat-

ing an adaptive model for environmental dynamics.

International Journal of Intelligent Information and

Database Systems, 5(3):201–228.

van Rijsbergen, K. (2013). The roots of the theoretical ba-

sis for information retrieval. In International Confer-

ence on the Theory of Information Retrieval, ICTIR

Utilizing Virtual Communities for Information Retrieval and User Modeling

’13, Copenhagen, Denmark, September 29 - October

02, 2013, page 19.

Wen, J.-R., Lao, N., and Ma, W.-Y. (2004). Probabilistic

model for contextual retrieval. In Proceedings of the

27th annual international ACM SIGIR conference on

Research and development in information retrieval,

pages 57–63. ACM.

Yakoubi, Z. and Kanawati, R. (2013). Leader-driven ap-

proach for community detection in complex network.

In proceedings of the international conference on in-

tercations in complex systems.

Zhang, D., Song, T., Li, J., and Liu, Q. (2014). A linked

data-based framework for personalized services infor-

mation retrieval in smart city. In Web Information Sys-

tems Engineering–WISE 2013 Workshops, pages 461–

473. Springer.

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies