MultiPerG: Multiple Periodic Geography convolution for next POI recommendation

Abstract Next Point-of-Interest (POI) recommendation aims to provide intelligent decisions about the next locations for users, according to their trajectories in location-based social networks. Existing next POI recommendation techniques profile users’ trajectories as sequences. However, most of them only consider consecutive correlations on sequences that are susceptible to time bias, while neglect regularity like daily repetition for periodic patterns, and may be corrupted by noise. Therefore, trajectories are explored insufficiently. To overcome the above limitations, we establish a Multiple Periodic Geography convolution network (MultiPerG) for next POI recommendation. MultiPerG is a hierarchical framework based on temporal convolutional networks. To capture both the multi-interval/span periodic patterns on sequences and the multi-scale geographical patterns on geography, three kinds of variant blocks are proposed as basic elements. Specifically, binary-tree/segment-tree deformable blocks are designed to perceive time periods in various intervals/spans and against bias. Geo-span deformable block is devised to capture spatial proximity in diverse scales, and alleviate the adverse effects of noise information. Extensive experiments demonstrate that MultiPerG outperforms the state-of-the-art next POI recommendation models. The ability of MultiPerG on exploring periodic patterns is also verified.