The Zen of Multidisciplinary Team Recommendation

In order to accomplish complex tasks, it is often necessary to compose a team consisting of experts with diverse competencies. However, for proper functioning, it is also preferable that a team be socially cohesive. A team recommendation system, which facilitates the search for potential team members can be of great help both for (i) individuals who need to seek out collaborators and (ii) managers who need to build a team for some specific tasks. A decision support system which readily helps summarize such metrics, and possibly rank the teams in a personalized manner according to the end users’ preferences, can be a great tool to navigate what would otherwise be an information avalanche. In this work we present a general framework of how to compose such subsystems together to build a composite team recommendation system, and instantiate it for a case study of academic teams.

💡 Research Summary

The paper presents a comprehensive framework for multidisciplinary team recommendation, called SWAT (Social Web Application for Team Recommendation), which integrates three orthogonal dimensions—competence, social cohesion, and collaboration history—into a unified graph‑based model. The authors argue that a “best” team cannot be defined by a single metric; instead, multiple quantitative measures must be considered. To this end they define:

1. Competence Graph – a labeled bipartite graph (I, EA, E) linking individuals (I) to expertise areas (EA). Each edge (i, ea) carries a real‑valued competence score c∈(0,1) indicating how proficient the individual is in that domain.

2. Social Graph – a directed multigraph (I, SE) where each edge is labeled with a tuple (d, s). The dimension d∈SD denotes the type of social relation (colleague, friend, mentor, etc.) and s∈(0,1) quantifies its strength. Multiple edges between the same pair of individuals are allowed to capture different dimensions, and directionality permits asymmetric strengths.

3. History Graph – an undirected bipartite hypergraph (I, EA, T) that records past collaborations. Each hyper‑edge t∈T is a pair (It, EAt) where It⊆I (|It|≥2) and EAt⊆EA (|EAt|≥|It|). This structure captures which individuals worked together and which expertise areas were covered, allowing the system to trace evolution of skills and team performance over time.

The framework is deliberately modular: data acquisition, entity extraction, graph construction, and recommendation ranking are separate subsystems that can be swapped with alternative techniques. The authors illustrate the instantiation of SWAT for academic research teams. Data are harvested from a variety of sources:

• Structured university staff records (NTU) for a clean initial corpus.
• Public scholarly databases (DBLP, Microsoft Academic, Google Scholar) for author lists, publications, venues, and citation counts.
• Academic social networks (Academia.edu, Facebook) to enrich the social graph with real‑world connections.
• Wikipedia category hierarchies to infer subsumption, similarity, and synonym relations among expertise areas.

Topic extraction from paper titles and abstracts is performed using a recent Wikipedia‑driven disambiguation method, providing a domain‑agnostic way to map publications to expertise concepts. The system also leverages Google Maps API to attach geographic metadata (country, region, city) to affiliations, enabling location‑aware team formation.

Because the raw crawled data are noisy, incomplete, and subject to temporal drift, the authors incorporate a crowdsourcing layer that lets end‑users correct competence labels, add missing relations, and validate inferred expertise. This feedback loop continuously improves the knowledge base.

The resulting knowledge base (as of the writing) contains roughly 996 k individuals, 553 k expertise concepts, and 2.94 M recorded teams, with an average of 7.27 social connections per person and an average team size of 3.12 members. Statistical analysis confirms the well‑documented trend toward larger, more collaborative research teams.

SWAT is delivered both as a standalone web application and as a Facebook app. Users specify a task by selecting required expertise areas and optionally adjusting social preferences (e.g., prioritize strong colleague ties, limit geographic spread, or enforce diversity). The system then enumerates candidate teams, computes three primary scores—competence coverage, social cohesion, and historical success (e.g., citation impact)—and aggregates them according to user‑defined weights. The ranked list is presented with visualizations of the underlying graphs, allowing users to explore why a particular team was suggested.

The authors emphasize that the architecture is domain‑agnostic: by redefining the sets I, EA, and SD and by plugging in appropriate data sources, the same framework can support corporate project teams, healthcare multidisciplinary groups, or educational collaborations. Future work is outlined to include systematic evaluation of recommendation accuracy, real‑time graph updates, and predictive models of team performance. Overall, the paper contributes a solid theoretical model, a practical implementation pipeline, and a demonstrable large‑scale system that bridges the gap between expertise discovery and socially coherent team formation.