Who is your friend: inferring cross-regional friendship from mobility profiles

Lingfei Ren, Ruimin Hu, Dengshi Li, Zheng Wang, Junhang Wu, Xixi Li, Wenyi Hu

Abstract: Location Based Social Networks (LBSNs) have been widely used as a primary data source to study friendship inference. Traditional approaches mainly focused on exploring pairwise co-location frequency, that it, the more frequency two users co-location, the more likely that they are friends. Such methods fail to solve the geographically restricted friends recommendation. In this paper, we tackle a novel friendship inference problem: cross-regional friendship inference, i.e., inferring whether users from different regions are friends. By revisiting mobility and social friendship of cross-regional friends based on a large-scale LBSNs dataset, we spot that cross-regional users are likely to form friendship when their mobility profiles are of high similarity. To this end, we propose Category-Aware Heterogeneous Graph Embedding Framework (CHGE) for inferring cross-regional friendship. We first utilize multi-bipartite graph embedding to capture users’ mobility neighbor proximity and activity category preference simultaneously, then the contributions of Point of Interest (POI) and category are learned by a category-aware heterogeneous graph attention network in an unsupervised method. Extensive evaluations on the real-world LBSNs dataset show that our CHGE significantly outperforms the state-of-the-art approaches by up to 9.7% on Area Under the ROC Curve (AUC) and 7.3% on Average Precision (AP).