Spatio-Temporal Position-Extended and Gated-Deep Network for Next POI Recommendation

Abstract

The next point of interest (POI) recommendation uses the user’s check-in information on location-based social networks to make recommendations. The existing methods based on deep learning are evident in improving the performance of the recommendation model by capturing users’ interests and preferences. However, the methods based on recurrent neural networks ignore the dependencies of non-continuous POIs for understanding users’ behaviour under spatio-temporal factors. Most attention-based methods focus on the global POI sequence, which pays attention to all POIs in the users’ check-in sequences, even if some attention has very little weight. To tackle these problems, we propose a novel spatio-temporal model based on the position-extended algorithm and gated-deep network (i.e., ST-PEGD) for next POI recommendation. Specifically, by combining spatio-temporal factors, we design a gated-deep network to capture the long-term behavioral dependencies of users by generating auxiliary binary gates. In addition, when capturing the short-term behaviour dependence of users, we use the position-extended algorithm to make the contextual interaction of RNNs more sufficient when performing POI sequence hopping selection. Extensive experiments on two real datasets prove that our model performs significantly better than state-of-the-art methods.