Fuzzy DA clustering-based improvement of probabilistic latent semantic analysis

Abstract

Probabilistic latent semantic analysis (pLSA) can be interpreted as a soft co-clustering model with an intrinsic fuzzification penalty and the partition quality was shown to be improved by tuning the degree of intrinsic partition fuzziness while the model is not supported by probabilistic constraints. In this paper, the mechanism of intrinsic fuzziness tuning is utilized for improving the partition quality of pLSA under the strict probabilistic constraints. The proposed deterministic annealing approach first initializes a co-cluster partition with a slightly fuzzier penalty weight and then gradually reduces the intrinsic fuzziness until it reaches the strict probabilistic constraints. Supported by the robust feature of fuzzier models against random initialization, the derived pLSA partition is demonstrated to be more stable in several numerical experiments.