Time-Sequential Variational Autoencoders for Recommendation

Abstract

In recent years, a variational auto-encoder (VAE) -based methods have been attracting attention in the study of recommendation systems and a VAE-based method extended to handle sequential information in order to consider the order of user's actions was proposed. However, this method only considers the order of actions, not the date and time of each action. If the date and time of the action can be incorporated into the information used for recommendation, since information based on time intervals between actions such as a product purchase interval and a user's maturity level for a product category can be incorporated, higher accuracy is expected. Therefore, we propose a VAE-based recommendation system that improves accuracy by adding the time information of each action to the input sequential information. We utilize Time-LSTM instead of GRU for RNN-encoder and it confirmed the improvement in recommendation accuracy.