Proposal of Initial Values Set for Estimating Parameters of probabilistic Latent Semantic Analysis

Abstract

Machine learning methods aimed at extracting rules from large amounts of data and discovering causal relationships are spreading in various fields. Many machine learning methods obtain parameters by iteratively reducing a loss function value from a reasonable initial value. Although there is an advantage that a solution can be obtained with the performance of a computer even if there are many parameters, there is also a problem that the obtained solution differs depending on the initial values to be set. In this paper, we propose a method of setting initial values for probabilistic Latent Semantic Analysis (pLSA), which is often used in marketing segmentation. This method not only obtains a unique solution by explicitly setting the initial value, but also aims to improve computational efficiency by reducing the number of iterations. We verify the proposed method using real data and confirm its effectiveness.