[3H5-GS-3] Data mining: Applied data mining (2)

[3H5-GS-3-01] Predicting user's next Point-of-Interest with GRU and attention mechanism considering time series variation of preferences and geographic features

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

Keywords:Point-of-Interest Prediction, Attention, spatiotemporal

Recently, researches on user's activity have been frequently conducted. Since human activities are caused by various factors and are difficult to predict, it is important capturing them in order to predict human activity.
In this study, we use GRU and attention mechanism to capture periodicity, short and long term transitions of preferences and geographic effects. With regard to POIs transitions, we capture periodicity and time series variations by GRU and use attention to the important history. Regarding geographic effects, we consider not only POIs, but features of regions users visited to capture the relationship between POIs and regions. In experiments, we predict whether a user will visit each POI category in the next day as an output taking sequences of visited POIs and regions as an input. Experiments using real-life datasets show that the proposed model outperforms the existing models.