Dynamic Topic-Enhanced Memory Networks: Time-series Behavior Prediction based on Changing Intrinsic Consciousnesses

Abstract

In the field of behavior prediction, methods have been developed to predict the state of the user by using the previous state or time-series of recorded behavior histories. However, so far, there has been no effort to capture time series reflecting the intrinsic consciousnesses and changes thereof of users. Here, we propose a model that captures changes in intrinsic consciousnesses of the user, called Dynamic Topic-Enhanced Memory Networks (DTEMN), for location-based advertising. In comparative experiments, we used DTEMN to predict places where users will visit in the future. The results show capturing changes in intrinsic consciousnesses using DTEMN is effective in improving prediction performance. In addition, we show an improvement in interpretability when simultaneously learning topics expressed as multiple intrinsic consciousnesses.