Deterministic annealing process for pLSA-induced fuzzy co-clustering and cluster splitting characteristics

Deterministic annealing (DA) is a powerful tool for escaping many poor local optima in fuzzy clustering. In this paper, a novel approach of improving the performance of probabilistic Latent Semantic Analysis (pLSA) is proposed supported by a DA process, where pLSA solutions are handled in the fuzzy co-clustering context. Although pLSA is defined under a purely statistics concept, it includes an intrinsic soft partition principle, which has a similar form to the entropy-based fuzzification in fuzzy c-means clustering. The proposed DA process is realized by tuning the intrinsic fuzziness degree of pLSA and is expected to improve the initialization sensitivity of pLSA solutions. Additionally, the cluster splitting characteristic of DA is also useful in cluster number selection, where a sequence of cluster splittings produces a hierarchy of fuzzy clustering solutions. The characteristic features are demonstrated through several numerical experiments including both artificial data sets and real world benchmark data sets.