A Method for Analyzing Solution Diversity in Topic Models

A topic model is a statistical model for modeling high dimensional count data. Many different parameters (solutions) of a topic model can be obtained through a learning algorithm due to different initial conditions. This paper focuses on diversity of solutions. To utilize diversity of solutions, it is necessary to acquire distribution structure of them. Therefore, this paper proposes a novel method to define similarity (inner product) of solutions using normalized mutual information to analyze distribution of solutions. Experimental results for text data are presented to show the usefulness of the proposed method.