Evolutionary Clustering via Message Passing

When data are acquired at multiple points in time, evolutionary clustering can provide insights into cluster evolution and changes in cluster memberships while enabling performance superior to that obtained by independently clustering data collected at different time points. Existing evolutionary clustering methods typically require additional steps before and after the clustering stage to approximate the number of clusters or match them across time. In this paper we introduce evolutionary affinity propagation (EAP), an evolutionary clustering algorithm that groups points by passing messages on a factor graph. The EAP algorithm promotes temporal smoothness via factor nodes that link variable nodes across time, and introduces consensus nodes that enable cluster tracking and identification of cluster births and deaths. Akin to the conventional (static) affinity propagation, the EAP framework automatically detects the number of clusters. The effectiveness of EAP is demonstrated on several simulated and real world datasets.