[1I4-GS-2] Machine learning: Applied machine learning (1)

[1I4-GS-2-03] Dynamic Topic Memory Networks: Time-series Behavior Prediction Based on Transition of Intrinsic Preferences

〇Ryoko Nakamura¹, Aozora Inagaki¹, Ryo Osawa¹, Toshikazu Fukami², Isshu Munemasa², Tomohiro Takagi¹ (1. Meiji University, 2. CyberAgent, Inc.)

Keywords:dynamic memory networks, episodic memory, time-series, behavior prediction

We predict behavior of users in the future based on past their behavior histories in order to improve the performance of location-based advertising. In the field of behavior prediction, researches using external time-series behavior history of the user has been developed. However, researches capturing the time-series transition of intrinsic preferences of the user are inadequate. Therefore, we propose the model capturing the multiple transitions of intrinsic preferences of the user, Dynamic Topic Memory Networks (DTMN). In the experiments, we predict the places where users will visit in the future. We show the effectiveness of capturing the transition of intrinsic preferences using DTMN by improving performance in comparative experiments. Through experiments, we also show the importance of capturing multiple transitions of intrinsic preferences.