Method for Extraction of Purchase Behavior and Product Character Using Dynamic Topic Model

In this study, we focus on extraction of latent topic transition from POS data. POS analysis is conducted to obtain the frequent pattern of customer's behavior. The fundamental method for POS analysis is to conduct market basket analysis. By doing Market basket analysis, the sets of products that are often bought at the same time can be extracted. In market basket analysis, however, the effect of time series is not considered. We conducted the experiment based on two hypotheses. One is that each product has several topics. The other is that the proportion of each product on a topic changes as the period changes. To extract topics and their changes, we use Dynamic Topic Model (DTM), which is an extended model of Latent Dirichlet Allocation (LDA). Then we obtain the change of the topic-word distribution on each topic. Different topic has different characters, but it seems that there is a relationship between each topic. Therefore, we conduct correlation analysis to several items. From the result of visualization of product features vector of several items, we can obtain that each product has unique time-series change of product feature. This study is also conducted to reveal product features vector based on Customer Purchase Behavior. By using DTM, each basket is transformed into probability distribution vector, and we use this topic vectors as each basket's evaluation result of topic features. We divide POS data into 12 groups by purchasing time and create a heat map that indicates changes of topic proportion as time advances. By conducting this analysis, we can grasp customer behavior based on the topic vector space. These analysises reveals product features are created based on topic correlations and its change and customer behavior can be extracted as the change of topic proportion, so the results show that the presented method is promising in the extraction of products' features and customer behavior.