Attention-Based Dynamic Preference Model for Next Point-of-Interest Recommendation

摘要

Next Point-of-Interest (POI) recommendation is a core task for various location-based intelligent services. It aims to predict the POI a user tends to visit next given the user's check-in history. However, this task is not trivial because of two challenges: 1) complicated sequential patterns of user mobility and 2) dynamic evolution of user preferences. In this paper, we propose a novel model called Attention-based Dynamic Preference Model (ADPM) for next POI recommendation, which can capture dynamic long- and short-term preferences of users. In particular, we design a spatio-temporal self-attention network to explore complicated POI-POI transition relationships and capture short-term sequential patterns of users. Moreover, we utilize a user-specific attention to learn evolving long-term preferences of users based on their constantly updated check-ins. Finally, a feature gating fusion module is introduced to adaptively combine long- and short-term preferences. We conduct extensive experiments on two real-world datasets, and the results demonstrate the effectiveness of our model.