Verification of Probabilistic Latent Semantic Analysis Clustering Solution Stability and Proposal of Optimal Initial Values Setting Method

摘要

pLSA is a useful method to know the characteristics of customer or item in marketing. In this study, we proposed a method to set the initial values more efficiently than the existing method for the problem that the final solution depends on the initial values set in the EM algorithm used by pLSA to estimate the solutions. We focused on the dimensional compression and clustering that are the characteristics of pLSA, and thought that the stability of the solution of pLSA would be improved by reflecting it in the initial values. Therefore, first, we performed correspondence analysis and k-means cluster analysis on the original data to express the features of dimensional compression and clustering. Next, we compared the performance of the pLSA results with the initial values of the proposed method and the initial values of the conventional method using random numbers. As a result, it was shown that the proposed method also converges to the same log-likelihood as the conventional method, and that the proposed method is superior in terms of convergence speed and stability.