On Using Social Context to Model Information Retrieval and Collaboration in Scientific Research Community - Community-Built Databases: Research and Development - page 149

Table 6.2 Comparison of social importance measures
p @0.1

p @0.2 p @0.1 p @0.2

Betweeness 0.0363 0.0363 W-Betweeness 0.0374 0.0398

Closeness 0.0232 0.0191 W-Closeness 0.0214 0.0189

PageRank 0.0324 0.0299 W-PageRank 0.0225 0.0199

Authority 0.0389 0.0411 W-Authority 0.0398 0.0423

Hub 0.0516 0.0430 W-Hub 0.0516 0.0433

The weighted model slightly improves the retrieval precision for most social
importance measures. This is approved with values obtained by W-Hub , W-Authority ,
and W-Betweeness measures beyond their analogous measures applied to a binary
social network. Therefore, we conclude that the properties expressed through
weights on social relationships including the shared interests, the influence, and
the knowledge transfer can better identify Central authors and then estimate the
relevance of bibliographic resources. For all the social importance measures, pre-
cisions p @0.1 and p @0.2 do not exceed the threshold of 60% compared with those
of the traditional information retrieval model based on the TF

IDF metric having

p @0.1 0.0786. Therefore, the social importance measures

are not able to sort the results without taking into account the similarity between
document and query.
In the remaining experiments, we retained the W-Hub measure as it is the best
measure expressing the social importance of bibliographic resources.

¼ 0.08 and p @0.2 ¼

6.5.3.3 Evaluation of Our Model Effectiveness

In order to evaluate the effectiveness of our model, we first select the best tuning
parameter a , then we compare the retrieval performances with similar retrieval

systems.

Tuning the Parameter a

We studied the impact of the parameter a on the retrieval process [see ( 6.8 )]. We

note that if

a ¼

0 only the social relevance is taken into account. Moreover,

a ¼

1

corresponds to the baseline TF IDF since the topical relevance is only considered

to rank documents.
We note in Fig. 6.4 a significant improvement in performance following the
integration of topical relevance with a value of

over 0.4. Analyzing precisions

a

p @0.1 and p @0.2 in function of the parameter , we note that curves present peaks

a

with values exceed obtained value for 0 and this is when only the topic

relevance is taken into consideration. Hence, the combination of the two scores
can effectively improve the final ranking of documents. The best values of the
parameter

a ¼

are obtained between 0.5 and 0.6.

a