A Study on Improving the Interpretability of Biterm Topic Model by Learning of Emphasized Data Augmentation

Abstract

The major users of EC sites have a small number of purchases are called light users. When we apply the topic model such as Latent Dirichlet Allocation for light users purchase history, the estimation accuracy is reduced because of the small amount of data. Therefore, Biterm Topic Model (BTM) has been proposed. BTM is a model that assumes the same topic for a pair of two items (biterm) in the user's purchase history. It learns topics with an emphasis on pairs of two items with a high simultaneous purchase probability. In general, the possibilities of simultaneously purchasing two items that are popular by many users tend to be high. However, even if the simultaneous purchase probability of a pair of items A and B is relatively low, the biterm with a high conditional purchase probability of item B conditioned by item A is more important for business purposes such as marketing policy. Therefore, in this study, we define pairs of items with high conditional purchase probabilities as important related biterms. We propose a new learning method by emphasizing these biterms. In addition, we apply our proposal for both artificial data and actual data to verify the effectiveness.