Weighted Evidence Accumulation Clustering Using Subsampling
F. Jorge F. Duarte, Ana L. N. Fred, Fátima Rodrigues, João Duarte, André Lourenço · 2006
We introduce an approach based on evidence accumulation (EAC) for combining partitions in a clustering ensemble. EAC uses a voting mechanism to produce a co-association matrix based on the pairwise associations obtained from N partitions and where each partition has equal weight in the combination process. By applying a clustering algorithm to this co-association matrix we obtain the final data partition. In this paper we propose a clustering ensemble combination approach that uses subsampling and that weights differently the partitions (WEACS). We use two ways of weighting each partition: SWEACS, using a single validation index, and JWEACS, using a committee of indices. We compare combination results with the EAC technique and the HGPA, MCLA and CSPA methods by Strehl and Gosh using subsampling, and conclude that the WEACS approaches generally obtain better results. As a complementary step to the WEACS approach, we combine all the final data partitions produced by the different variations of the method and use the Ward Link algorithm to obtain the final data partition.