Matching Resources in Social Environment
Amel Benna
1,2
, Hakima Mellah
1
, Islam Choui
3
and Ali Oualid
3
1
CERIST, 05, Rue des 03 Fr`eres Aissiou, BP 143, BenAknoun, 16030 Algiers, Algeria
2
USTHB, BP 32, El-Alia Bab-Ezzouar, Algiers, 16111, Algeria
3
ESI, BP 68M, Oued Smar, 16309, Algiers, Algeria
Abstract. User comments on the web are becoming more and more important.
We focus, in this paper, on the use of user-defined tags for annotating resources
to identify links between them. These links are based on a social context of the
resource, obtained by applying k-means classification method and a hierarchi-
cal classification of tags within a cluster. The resources are re-assigned to this
classification to facilitate the search process. The ranking of results is performed
according to their degree of relevance, by evaluating a similarity score between
the tagged contents, in hierarchical clusters of tags, and the user request. The re-
sults of the evaluation, on the social bookmarking system del.icio.us, demonstrate
significant improvements over traditional approaches.
1 Introduction
User experience and comments on the web are becoming more and more important.
In 2010, Gartner group, predict that within five years, 70 percent of collaboration and
communications applications designed on PCs will be modelled after user experience
lessons from smart-phone collaboration applications
4
.
A collaborative tagging system put users at the centre of data production and intro-
duces a strong social collaboration. It describes the process by which many users add
meta-data in the form of keywords to shared contents [1]. These keywords require no
skill from user and are named tags. They and can be associated with different types of
resources (videos, images, bookmarks, articles, application and blogs).
The analysis of collaborative tagging systems structure showed regularities in user
activity, tag frequencies, kinds of tags used and a remarkable stability in the relativepro-
portions of tags within a given resource. Empirically, once a resource has been tagged
over a hundred times, each tag’s frequency, in a proportion, remains stable compared to
the total frequency of all other tags used for this resource [1]. However, works on link-
age information often do not take into account social information of resource that can
be retrieving from users significant tags. Indeed, the Social Information Retrieval (SIR)
follows from its domain model [2] and the incorporation of social factors can increase
the relevance of results returned in information retrieval [3],[4],[5],[6].
4
http://info.absnt.com/
Benna A., Mellah H., Choui I. and Oualid A..
Matching Resources in Social Environment.
DOI: 10.5220/0004088800610070
In Proceedings of the 10th International Workshop on Modelling, Simulation, Verification and Validation of Enterprise Information Systems and 1st
International Workshop on Web Intelligence (WEBI-2012), pages 61-70
ISBN: 978-989-8565-14-3
Copyright
c
 2012 SCITEPRESS (Science and Technology Publications, Lda.)

Our interest is focused on the use of user-defined tags for annotating resources to
identify links between resources. These links are based on a social context of the re-
source in folksonomy. A folksonomy is a system of classification derived from the
practice and method of collaboratively creating and managing tags to annotate and cat-
egorize contents [7]. A resource social context is related to purified and classified tags,
obtained by a classification method, and the words content refers to a resource or por-
tion of resource.
More specifically, we propose an approach for SIR that integrates social relation-
ships between contents by taking into account the social information of resource. A
social linking between resources is based on clusters. These latter are, a set of seman-
tics links between users purified tags, obtained by a classification method. We operate
the most significant tags in research. E.g. The adjectives tags such as ”funny”, “inter-
esting”, “mydocument” or words they do not even exist in the literature as ”xfd4” are
eliminated for low frequency use. We reduce redundancy or ambiguity of tags by find-
ing semantically related tags as tags from the same cluster. The proposed process for
modelling links between resources classifies and structures a folksonomy and includes
it in the matching and the ranking of search results.
The rest of this paper is structured as follows: Section 2 introduces some related
works on SIR based on linkage information and collaborative tagging systems. Section
3 proposes our model for linking resources in information retrieval based on collabora-
tive tagging. Section 4 presents the results of our evaluation. We conclude and highlight
future research direction in section 5.
2 Related Work
Several approaches have been proposed for using linkage information on the Infor-
mation Retrieval (IR) systems. This work can be distinguished according to different
factors (content, HTML, architecture, links, social, trust, personal)
5
.
The most famous proposed approaches for using linkage information, to aid in rel-
evant document retrieval, are PageRank [9] and HITS [10]. PageRank technique is cal-
culated independent of any query but pages with high PageRank are highly ranked
even though they are not relevant to a user’s query. Unlike PageRank, HITS is a query-
dependent form of linkage analysis. Two scores; authority and Hub are calculated for
each document. Evaluation of model [11], based on the works [2],[12], proves that the
extent of Hub (the centrality of the authors) is the measure to better assess the social
significance of documents. However, if the initial query expressed by a user does not
cover a sufficiently broad topic, there will often not be enough relevant pages. The main
disadvantage of this approach is that not only requires extra resources from the search
system at query time but also increases the system response time. The model proposed
in [17] illustrates an example of a study by applying four centrality measures (degree,
PageRank, closeness and betweenness) to evolving co-authorshipnetwork. In this work,
the measures of centrality include the impact of the resource, i.e. its citing accounts and
scope of author.
5
From Periodic Table of Search Engine Optimisation Ranking Factors
(http://SELND.COM/SEOTABLE)
62

Search engines such as Google, Yahoo and Bing use several factors to retrieve in-
formation, some factors may influence more than others and may be considered more
important than others. However, no single factor guarantees a relevant research and top
rankings
4
.
To improve the web search, various approaches [4],[5],[13]explore the use of social
annotations. In [4] two new algorithms are proposed: the first one calculates the simi-
larity between social annotations and web queries whereas the second captures the pop-
ularity of web pages using social annotations. A model in [5], based on social approval
votes of documents, shows that social information on documents can improve research
and the approval sources provide more details on user needs, particularly, when votes
are provided by experts. To define user expertise level, a user model in [13] is integrated
in calculating the tag weight. The evaluation is based on the closeness degree between
user interest’s and resource area, expertise and personal assessment for tags associated
to the resource.
Nevertheless, IR systems that use collaborative tagging suffer from a number of
limitations such as: variability on writing some tags, ambiguity due to the existence of
synonyms, an the absence of semantic links between tags. These leads to impoverish
information research potential whereas the rate of tagged contents is growing every day,
and affect the response time and the result quality. Data clustering has been used, for
natural classification, to identify the degree of similarity among forms or organisms,
and for compression, as a method for organizing the data and summarizing it through
cluster prototypes. A cluster of tags represents the most common way to gather addi-
tional information in collaborative tagging systems [8]. It was defined to:
- Use the most significant tags [14],
- Decrease redundancy or tags ambiguity [15],
- Find the similar semantic tags [14],
- Reduce the response time and improve the quality of results[16]. Thousands of clus-
tering algorithms have been proposed in the literature. Nevertheless, clustering meth-
ods differ on the choice of the objective function, probabilistic generative models, and
heuristics [8]. The K-means [8] method is used to classify tags of folksonomies such as:
customizing folksonomies based clustering of tags [14], extraction of relationships be-
tween users and resources tagged based clusters of tags [18]. The experience on Word-
net ontology, in [19], showed that the tags associated through simple co-occurrence
measures tend to maintain subsumption relationships (a hierarchical relationship be-
tween concepts), whereas tags associated via a similarity distributional measure in the
context tag-tag tend to be at the same hierarchical level, or to share the same parent/
grandparent.
The works on linkage information we have cited do not take into account social
information of resources, which can be retrieved from significant users tags; The social
factor focuses only on a social reputation of a user account and on user social shares
in social network and neglect social information of resources and links between tags.
Inspired by this works and to consider social factors for using linkage information,
we modelled links between resources for social research, using a purified tags and a
hierarchical structure of cluster of tags.
63

3 Social Linking Model
The social linking model, based on practices of collaborative tagging, is used to define
links between resources (see Fig. 1). The definition and the structure of the proposed
folksonomy for linking resources are presented in section 3.1. We describe the social
search process that explores folksonomy in section 3.2 and the evaluation of results in
section 3.3.
Fig.1. Social linking resources architecture.
3.1 Folksonomy Structure
Let consider F=(U,T,R,Y) the formal structure of a folksonomy [22]. U, T and R are
finite sets, whose elements are respectively users, tags and resources. Y is a ternary
relation between them such that:
Y ⊂ U × T × R
A post is a triple (u, t
u
r
, r), where, t
u
r
∈ T is a tag used by user u, u ∈ U to tag a
resource r, r ∈ R . The classification process of our folksonomy includes four steps:
1. Creating a semantic tag-tag data matrix,
2. Generating clusters of tags,
3. Defining hierarchy of tags within a cluster,
4. Assigning resources to cluster of tags hierarchy.
Semantic Tag-tag Data Matrix. To define a link between tags we compute the co-
occurrence matrix in the context tag-tag [19]. This co-occurrence is determined by the
co-occurrence W(t
i
, t
j
) between each pair of tags (t
i
, t
j
) as in (1).
W (t
i
, t
j
) = |(u, r) ∈ U × R/(u, t
i
, r) ∈ Y (u, t
j
, r) ∈ Y | (1)
64

The data matrix is then transformed into a cosine matrix [20] by measuring the
cosine distance between vectors, as in (2), where a vector,
−→
v
t
i
, denotes the number of
times a user U
i
uses a tag t
j
and it is computed as described in [23].
cos(
−→
v
t
i
,
−→
v
t
j
) =
−→
v
t
i
·
−→
v
t
j
k
−→
v
t
i
k · k
−→
v
t
j
k
(2)
Generating Clusters of Tags. To minimizing within clusters variance tags and maxi-
mizing the distance between clusters of tags, we apply the k-means method on the co-
sine matrix. The K-means algorithm requires three user-specified parameters: number
of clusters K, cluster initialization, and distance metric. The most critical choice is K.
Whereas no perfect mathematical criterion exists; a number of heuristics are available
for choosing K. One way to overcome the local minima is to run the K-means algo-
rithm, for a given K, with multiple different initial partitions and choose the partition
with the smallest squared error.[8] After applying k-means to generate set of k-clusters
of tags, we use the Levenshtein distance [21] to avoid spelling variations of tags and
composite words in each cluster of tags.
Cluster of Tags Hierarchy. A hierarchy of tags in each cluster is build. Each tag,
with its variant spellings are grouped into a single concept by applying hierarchical
classification algorithm [22]. This hierarchy structures the clusters of tags as a tree,
where tags are tree nodes and resources tree leafs. We design by tags path, any path
leads from the root node (the most common tag used in the cluster) to a leaf node (tags
used less in the cluster). The tag that has a high degree of co-occurrence in the resources
is chosen as a concept.
Assigning the Resources to Cluster of Tags Hierarchy. In order to form clusters that
contain similar resources, the resources tagged are reassigned first to clusters of tags.
A resource, r
i
, degree of membership, D
r
i
c
j
, to the cluster c
j
is computed as in (3).
occ(t
l
, r
i
) denotes co-occurrence of tag t
l
with a resource r
i
and t
l
belongs to cluster
c
j
.
D
r
i
c
j
=
P
t
l
∈c
j
occ(t
l
, r
i
)
|u ∈ U/(u, t, r
i
∈ Y )|
(3)
To determine resource tags in cluster of tags, each resource, r
i
, is associated to the
hierarchical cluster of tags whose degree of belonging to it is maximal.
After having classified the folksonomy F into clusters of tags, defining hierarchy of
tags in each cluster, and reassigning resources to tags, an XML file is used to store the
structuring folksonomy, i.e. hierarchy of clusters tags, tags and associated resources.
3.2 Social Search Process
To answer a user query, the first step of the social search process is the query expansion,
the second one is the matching between clusters of tags and request tags and the last
step is the ranking of results.
65

Query Expansion. When a user issues a query, it is disambiguated by detecting vari-
ations in spelling of its keywords. The Levenshtein distance is used with a threshold
equal to 0.8. Indeed, most tags are names, and thus the lemmatization methods are not
recommended [15]. After query disambiguation, a linguistic ontology is used to deter-
mine semantic of request tags. In fact, request keywords are considered as tags. The
objective is to guide the user by suggesting keywords related to the meaning of the re-
quest word.
E.g. When the WordNet ontology is used for the word Java, three senses are proposed:
island, coffee, object-oriented programming language. The user request is enriched by
tags language, object- oriented, and programming for computer science user interest
area.
Matching and Ranking Resources. To answer a user find resources, we first identify
clusters of tags that match with the user request tags, and then search for tagged content
matching user query. Because tags are structured in hierarchical cluster, the user query
tags can match the cluster of tags tree nodes. As tags are close together, there is great
probability that request tags belong to the same cluster of tags. To identify clusters of
tags matching user request, a semantic similarity score, Jaccar d(
−→
V
r
,
−→
V
c
i
), is computed
between each vectors of clusters,
−→
V
r
, and query expansion vector,
−→
V
c
i
, as in (5). The user
request is represented by a tags vector,
−→
V
r
, and each cluster of tags, c
i
, is represented by
tag vector
−→
V
c
i
.
Jaccard(
−→
V
r
,
−→
V
c
i
) =
−→
V
r
⊓
−→
V
c
i
k
−→
V
r
⊔
−→
V
c
i
k
(4)
After identifying cluster of tags, identifying resources that meet user request means
to browse tree, of the selected clusters of tags, looking for tree leaves in which nodes
match the query tags. To determine such leaves, for each cluster of tags whose tags
match the tags of the query we proceed as follows:
1. For each node, we select all the related contents where nodes tags match user request
tags.
2. If in the same path, there is more than one tag that matches user request tags, we
select the deepest one in the subtree.
E.g. let language, object, java, javascript be tags of user expansion request for query
keyword Java. The tags Object and Java are in the same tree path (see fig. 2), but the
tag Java is deeper than the tag Object in the tags hierarchy. The content C
T
3
tagged by
jsp is select as request result. The contents C
T
1
, C
T
2
are also selected for the javascript
tag, as leaf of node javascript in cluster of tags subtree.
Results Ranking. The responses to a query may be found in a single content or may be
subject to an aggregation of a set of results shared with different contents. The aggre-
gation of contents, returned by query results, is to combine contents that match the user
request but that was tagged by different users of the system. This aggregation includes
any type of resource (text document, image and video). For the example in Fig. 2 the
contents C
T
1
, C
T
2
and C
T
3
are aggregated and are displayed to end user as one result.
The ranking of resources returned in a search is performed according to their degree of
66

Fig.2. Example of hierarchy of tags
relevance. We compute similarity score, Sim(
−−→
V
t
c
i
,
−→
V
r
), between request and content by
using the, most commonly used measure, cosine of the angle between the query vector,
and tagged content vectors. This score is computed as in (6). T
k
denotes the k
th
tags of
request.
Sim(
−−→
V
t
c
i
,
−→
V
r
) =
P
t
k
∈
−→
V
r
occ(t
k
,
−−→
V
t
c
i
)
|u ∈ U/(u, t, r
i
∈ Y )|
(5)
3.3 Evaluation and Results
To evaluated our approach, we extracted data from ’delicious data’ which contains a set
U of 2000 users, a set T of 2000 tags, and a set R of 70 resources for 3577 annotations.
Data analysis for 2000 tags showed that 200 tags have a high co-occurrence frequency
for 1879 users and represent more than 70% of users annotations. We had used only the
triplets of tags, users and resources. These triplets represent 80% of folksonomy tags.
As a first step, we seek to build clusters of tags (see Fig. 3). The set T of tags has
been classified using k-means method, with k = 17. The result is a set C of clusters with
an average of 12.3 tags for each cluster. The similarity distance between two resources
assigned to a cluster is greater than 0.65. The resources that have a similarity degree
more than 0.9 are grouped in the same slice of branch in the hierarchy To Define a
hierarchy of the ordered list in cluster of tags,we measure the cosine similarity between
vectors for λ > 0.5.
To evaluate the relevance of our approach, a series of tests for two kind of research
are performed: a traditional IR, based on the vector model, and a SIR, based on the
model that we define in section 3. Fig. 4 illustrates an example of the relevance mea-
sure, for the top 5 recommended resources, for a query Java in the music interest area.
The recall-precision curves measures vary inversely, precision decreases as the recall
increases. We observed that SIR search performs better than traditional IR search.
67

Fig.3. Dendrogram of similarity between tags.
Fig.4. Example of recall and precision evaluation for query.
4 Conclusions
We describe the social linking model based on practices of collaborative tagging to
define links between resources. The structure of the proposed folksonomy allows as-
signing resources to hierarchical structure of tags within cluster. We apply the cluster-
ing algorithm, k-means, to define clusters of tags and hierarchy in each cluster. A user
query can be expanded, by suggesting tags from the cluster of tags and ontology. In
our social search process we define a similarity degree for ranking function to classify
aggregated contents according to their relevance degree.
The evaluation was flown on the bookmarking system del.icio.us. The first conclusions
that emerge from evaluation of relevance tests are that the similarity distance between
resources within cluster of tags are very close and the integration of resource social
context provides very conclusive over traditional approaches. Our future works are ori-
68

ented towards the definition of local ontology from the hierarchy of tags within clusters
of tags.
References
1. Golder, S., Huberman, A. B.: The Structure of Collaborative Tagging Systems, CoRR
abs/cs/0508082. 18 August 2005.
2. Kirsch, S.M., Melanie, G., Cremers B. A.: Beyond the Web: Retrieval in Social Informa-
tion Spaces,.In Advances in Information Retrieval. London, UK : Springer, Lecture Notes in
Computer Science,2006, Vol. 3936, 84-95.
3. Zanardi,V., Capra,L.: Social Ranking: Uncovering Relevant Content Using Tag-based Rec-
ommender Systems, RecSys’08, 23-25 October 2008.
4. Bao, S. , Gui-Rong, X., Xiaoyuan,W., Yong, Y., Fei, B., Su, Z.: Optimizing web search using
social annotations, In Proceedings of the 16th Internationa Conference on World Wide Web,
WWW 2007. ACM 2007, 8-12 May 2007, 501-510.
5. Kazai, G., Milic-Frayling, N.: Effects of Social Approval Votes on Search Performance, In
ITNG 2009, Sixth International Conference on Information Technology: New Generations,
27-29 April 2009, ISBN 978-0-7695-3596-8, 1554-1559.
6. Benna, A., Mellah, H., Hadjari, K.: Building a social network, based on collaborative tag-
ging, to enhance social information retrieval, In ICITES, 2012, 453-458.
7. Isabella, P.: Folksonomies. Indexing and Retrieval in Web 2.0, K G Saur Verlag, 2009.
8. Jain, A.:Data Clustering: 50 Years Beyond K-Means, 2009.In Pattern Recognition Letters,
2009.
9. Page, L., Brin, S., Motwani, R., Winograd, T.:The PageRank citation ranking: Bringing order
to the web, In WWW98. 1998, 161172.
10. Kleinberg, J.M.: Authoritative Sources in a Hyperlinked Environment,In SODA, ACM,
1999, Vol.46, 668-677.
11. Ben Jabeur, L., Tamine, L., Boughanem, M.: A social model for literature access: towards a
weighted social network of authors,RIAO 2010, 32-39.
12. Mutschke,P.: Enhancing Information Retrieval in Federated Bibliographic Data Sources Us-
ing Author Network Based Stratagems, In ECDL 2001, LNCS 2163, pp. Springer, 4-9
September, Vol. 2163, 287-299.
13. Kichou, S. , Mellah, H., Amghar, Y., Dahak F.: Tags Weighting Based on User Profile, In Ac-
tive Media Technology - 7th International Conference, AMT 2011. Lecture Notes in Com-
puter Science 6890 Springer 2011, 7-9 September 2011, ISBN 978-3-642-23619-8, 206-216.
14. Gemmell, J., Shepitsen, A., Mobasher, B., Burke, R.D.: Personalizing Navigation in Folk-
sonomies Using Hierarchical Tag Clustering, In DaWaK 2008, 196-205.
15. Spiteri, L.: Structure and form of folksonomy tags: The road to the public library catalogue,
2007, Vol. 4.
16. Begelman,G., Keller P., Smadja, F.: Automated tag clustering: Improving search and explo-
ration in the tag space, Proc. of the Collaborative Web Tagging Workshop at WWW., May
2006, 2226.
17. Yan, E., Ding, Y.: Applying centrality measures to impact analysis: A coauthorship network
analysis, 2009.
18. Grahl, M., Hotho, A., Stumme, G. : Conceptual Clustering of Social Bookmarking Sites, In
LWA. Workshop. September 2007, 50-54.
19. Cattuto, C., Ben, D., Hotho, A., Stumme, G. : Semantic Grounding of Tag Relatedness in So-
cial Bookmarking Systems, In The Semantic Web - ISWC 2008, 7th International Semantic
Web Conference, 26-30 october, 2008, 615-631.
69

20. Cattuto, C., Loreto, V., Pietronero, L. : Collaborative Tagging and Semiotic Dynamics,In
CoRR. May 2006.
21. Levenshtein, V. : Binary codes capable of correcting deletions, insertions, 1966.
22. Hsieh, W-T., Lai, W-S., Chou, S-C. : A collaborative tagging system for learning resources
sharing, In IV International Conference on Multimedia and Information and Communication
Technologies in Education (m-ICTE2006), 2006, 1364-1368.
23. Markines, B., Cattuto, C., Menczer, F., Benz,D., Hotho, A., Stumme, G.: Evaluating similar-
ity measures for emergent semantics of social tagging. In WWW 2009, 2009, 641-650.
70