Predicting Multi-actor collaborations using Hypergraphs

Social networks are now ubiquitous and most of them contain interactions involving multiple actors (groups) like author collaborations, teams or emails in an organizations, etc. Hypergraphs are natural structures to effectively capture multi-actor interactions which conventional dyadic graphs fail to capture. In this work the problem of predicting collaborations is addressed while modeling the collaboration network as a hypergraph network. The problem of predicting future multi-actor collaboration is mapped to hyperedge prediction problem. Given that the higher order edge prediction is an inherently hard problem, in this work we restrict to the task of predicting edges (collaborations) that have already been observed in past. In this work, we propose a novel use of hyperincidence temporal tensors to capture time varying hypergraphs and provides a tensor decomposition based prediction algorithm. We quantitatively compare the performance of the hypergraphs based approach with the conventional dyadic graph based approach. Our hypothesis that hypergraphs preserve the information that simple graphs destroy is corroborated by experiments using author collaboration network from the DBLP dataset. Our results demonstrate the strength of hypergraph based approach to predict higher order collaborations (size>4) which is very difficult using dyadic graph based approach. Moreover, while predicting collaborations of size>2 hypergraphs in most cases provide better results with an average increase of approx. 45% in F-Score for different sizes = {3,4,5,6,7}.

💡 Research Summary

The paper tackles the problem of forecasting future collaborations that involve more than two participants—such as research teams, project groups, or email threads—by representing the underlying social network as a hypergraph rather than a conventional dyadic graph. In a hypergraph, a single hyperedge can connect an arbitrary number of vertices, thereby preserving the joint occurrence information of all members of a group. The authors focus on the “old‑edge prediction” task, i.e., predicting the re‑occurrence of hyperedges that have already been observed in the historical data, because predicting completely new hyperedges is considerably more challenging.

To capture temporal dynamics, the authors introduce a three‑dimensional hyper‑incidence tensor. The first dimension indexes hyperedges (or dyadic edges in the graph baseline), the second dimension indexes vertices (actors), and the third dimension indexes discrete time snapshots. For each snapshot of width w (e.g., a year or a five‑year window), they construct a hyper‑incidence matrix H(t) whose rows correspond to all distinct hyperedges seen up to that time and whose columns correspond to all actors. An entry H(t)