Tracking the evolution of communities in co-authorship networks: A semantically aware approach

This paper addresses the problem of tracking the time evolution of communities within co-authorship networks. We consider an evolutionary clustering approach which adapts the statistical framework of shrinkage estimation to obtain a smoothed version of the overall affinity matrix as the optimal weighted average between the matrices of past and current affinities. Moreover, the current affinity matrix at each time step is formed as the optimal mixing between the raw structural relationships between authors coupled with the semantic similarity of their published works, captured by performing LDA-based topic modeling on the corresponding corpora of abstracts. The proposed parametric weighting scheme is simultaneously optimized on a real co-authorship dataset emerging from the network of participants in the ICMB conferences, held from 2002 to 2013. Finally, community detection at each time step is conducted by employing spectral clustering on the estimated overall weight matrix. The obtained results justify that our approach provides a clearer revelation of the inherent authors´ communities dynamics when compared against incremental cluster formations that rely exclusively on the structural information of the co-authorship network.