Determining the best K for clustering transactional datasets: A coverage density-based approach

The problem of determining the optimal number of clusters is important but mysterious in cluster analysis. In this paper, we propose a novel method to find a set of candidate optimal number Ks of clusters in transactional datasets. Concretely, we propose Transactional-cluster-modes Dissimilarity based on the concept of coverage density as an intuitive transactional inter-cluster dissimilarity measure. Based on the above measure, an agglomerative hierarchical clustering algorithm is developed and the Merging Dissimilarity Indexes, which are generated in hierarchical cluster merging processes, are used to find the candidate optimal number Ks of clusters of transactional data. Our experimental results on both synthetic and real data show that the new method often effectively estimates the number of clusters of transactional data.