T-RecS, besides being a vehicle to carry out, validate and evaluate our multi-dimensional social network analysis driven expert and multidisciplinary team recommendation algorithms, is also an utility which can be used by, and in exploring NTU's researchers according to their expertise and collaboration (social) connections. It can be checked out using our Web demo (check the top-menu).

Features and screenshots

While our approach is general enough to incorporate other domains (subject to availability of necessary data, for example), the current implementation of T-RecS is an application leveraging on academic knowledge networks to recommend and help users explore multidisciplinary scientific teams. We have instantiated and deployed this prototype for local usage at NTU Singapore, using publicly available academics web sites and publications records. Although the case study is restricted to academic networks, and uses only public domain information, the T-RecS framework can be easily extended, possibly with addition of new elements to the modular architecture (see Figure below) for team recommendation in other (e.g., corporate) settings.

Query expression

The different skills for which experts are being sought can be queried. The main issue in this matter is that users are likely to use wrong/inaccurate terms. Our solution is to make suggestions to the user while she enters her query. This possible formulations are based on a set of heuristics using both wikipedia pages (articles and categories) as well as explicit keywords cited by academics on their web pasges and on NTU web research directory. T-RecS can hence find common upper categories which can express user needs in more general terms, common informative subcategories, etc.

In the example below, "web" is refined in various other research topics as "blogs", "Semantic Web", "Web Search Engines", etc. Likewise, a query on "Performance" would find "High Performance Computing" which is an important topic for NTU computer engineering scientists. More spectacular is probably "strategy auction" which is reformulated with "Game Theory" for this category is a generalization of the two former ones.

Expert identification for each skill set

A ranked list (based on competence score mined from the publication records and personal web pages) for each queried skill is created, and presented to the user. The same person may have expertise in the different keywords being queried.

It is possible (see Figure below) to refined the query by asking for only academics with specific status, belonging to specific school, etc.

Expert visualization and manipulation

On clicking on the corresponding button, a java applet pops up and displays the social network of an academic (1 hop distance only, ie direct neighbours). It is then possible to have a description of her neighbours, and to find papers they published together. This view helps to know how experts are integrated in NTU academics community.

Team recommendation

A list of teams (with different combinations of experts) which cover all the necessary/queried skill-sets is presented to the user, ranked according to a scoring function combining the team's competence score (derived from the competence score of the team members) and also the team's social cohesiveness (derived from social network analysis of the members). These two individual scores are combined to derive the composite score using an user defined weightage.

Several modifications are proposed to the user, like relative importance of each skill, cohesiveness/expertise trade-off, etc. You could play with that values to modify teams accordingly.

Team visualization and manipulation

Each of the teams in the list can further be visualized. The visualization superimposes a social network graph among the team members, and a bipartitie graph between these experts and the competences. The strength of the pair-wise relations in the social network graph are derived from a co-authorship based web-of-trust. This web-of-trust can be explored by right-clicking on the corresponding link. The edges of the bipartite graph conencting experts with expertise are labeled with the degree of competence for the topic, as derived using some standard Information Retrieval algorithms on the researchers' publication lists.


Acknowledgements & Disclaimers

T-RecS has been developed by researchers under Asst. Prof. Anwitaman Datta's supervision in the SANDS Group at NTU Singapore in the context of mTeam project funded by A*Star SERC Grant. We will also like to thank NTU's RSO (Research Support Office) for providing the NTU academic staff's publication lists, which while mostly publicly available, was easy to process in the structured format from RSO's RIMS database.
The expertise for each topic shown by the T-RecS system are output of algorithms. These algorithms, some of them obtained from the literature and some others developed by ourselves, are possibly far from perfect, and the research community as a whole continuously continues to refine them. Likewise, the data-sets we use are not complete, e.g., there are so many different kind of social relations people may have, which are not reflected by just co-authorship. Likewise, we do not take into account citations at the moment, though we intend to do so in the future, to determine expertise. Consequently, the scores - both on individuals' competence on a topic as well as social cohesiveness should not be considered as a reflection of the ground truth. In fact, one may say, there is no absolute ground truth for such subjects. Furthermore, the dataset we currently use is based on publications in 2007-8, and researchers' interest and profile may have changed in the meanwhile. The current web-demo should be treated strictly as a proof-of-concept.

For any further queries or feedback on T-RecS, contact