Pairwise Probabilistic Clustering Using Evidence Accumulation

Samuel Rota Bulò, André Lourenço, Ana L. N. Fred, Marcello Pelillo

Abstract: In this paper we propose a new approach for consensus clustering which is built upon the evidence accumulation framework. Our method takes the co-association matrix as the only input and produces a soft partition of the dataset, where each object is probabilistically assigned to a cluster, as output. Our method reduces the clustering problem to a polynomial optimization in probability domain, which is attacked by means of the Baum-Eagon inequality. Experiments on both synthetic and real benchmarks data, assess the effectiveness of our approach.