Recommending irregular regions using graph attentive networks

Abstract Due to the prevalence of human activity in urban spaces, recommending ROIs (region-of-interests) to users, especially irregular ROIs, becomes an important task in location-based social networks. A fundamental problem is how to aggregate users’ preferences over POIs (point-of-interests) to infer the users’ region-level mobility patterns. The majority of existing studies ignore the users’ implicit interactions with individual POIs when addressing this issue. For example, a user check-in a region cannot provide any specific information about how the user likes this region (we call this phenomenon “ROI-level” implicitness) and which POI in this region the user is interested in (i.e., “POI-level” implicitness). Furthermore, existing studies adopt predefined strategies for region-level preference aggregation, that is, initializing the importance of different POIs with identical weights, which is insufficient to model the reality of social networks. We emphasize two facts in this paper: (1) there simultaneously exists ROI-level and POI-level implicitness that blurs the users’ underlying preferences; and (2) individual POIs should have non-uniform weights and more importantly, the weights should vary across different users. To address these issues, we contribute a novel solution, namely GANR 2 (Graph Attentive Neural Network for Region Recommendation). Specifically, to learn the user preferences over irregular ROIs, we provide a principled neural network equipped with two attention modules: the POI-level attention module, to select the informative POIs of one ROI, and the ROI-level attention module, to learn the ROI preferences. Moreover, we learn the interactions between users and ROIs under the NGCF (Neural Graph Collaborative Filtering) framework. Extensive experiments on two real-world datasets demonstrate the effectiveness of the proposed framework.