A Bayesian nonparametric topic model for repeated measured data

Abstract

Topic models are currently used in many fields, particularly for marketing or medical science data analysis, often where an individual subject is repeatedly measured. A topic tracking model (TTM) that can consider the persistency of topics of individual subjects has been already proposed. Although the TTM estimates several parameters for each timepoint through online learning, offline learning should be utilized for analyses of preexisting data sets. Additionally, when a topic model is used, the number of topics should be decided in advance. However, deciding an appropriate number of topics is often difficult. Therefore, we propose a TTM with offline learning and a Bayesian nonparametric TTM (BNPTTM) for time-series data sets where data from individual subjects are repeated measures. The performance of the proposed topic model is evaluated using an actual prescription data set. Our results suggest that the TTM with offline learning has better predictive ability than the existing TTM, and the BNPTTM can deduce the number of topics from a given data set.