Dissimilarity Increments Distribution in the Evidence Accumulation Clustering Framework

摘要

In this paper, we combine two concepts. The first is the Evidence Accumulation Clustering framework, which uses a voting scheme to combine clustering ensembles and produce a co-association matrix. The second concept are Dissimilarity Increments, which are a high order dissimilarity measure which can identify sparse clusters, since it uses three data points at a time instead of two points, as in Euclidean distance. These two concepts are combined to form a new family of clustering algorithms, where the co-association matrix is used to form a distance which is then used to compute dissimilarity increments. These clustering algorithms are shown to improve the clustering results when compared to the usual Evidence Accumulation Clustering framework.