When Friendship Meets Sequential Human Check-ins: Inferring Social Circles with Variational Mobility

Abstract

The pervasiveness of mobile devices and the popularity of positioning technology generate diverse location-based social network platforms, which allow users to share POIs, post messages, and make online friends. Trajectory-based social circle inference (TSCI), aiming at inferring the social ties based on human mobility data, is one of the fundamental tasks that may facilitate a range of downstream applications such as group recommendation/advertising and crowdsourcing. Despite their promising results on TSCI, existing solutions still confront with three major challenges, including (1) inadequate representation learning ability of contextual information in the sparse check-in data; (2) the lack of a more generalized trajectory encoding mechanism for mobility pattern discovering; and (3) the neglect of modeling the explicit label correlations. To overcome these challenges, we propose a Graph-based Social Circle Inference (GSCI) framework to exploit implicit human mobility patterns and integrate the inherent correlations among social ties of users. We propose to integrate POI’s contextual information into a density representation from the perspective of graph learning rather than solely relying on the sequential visiting behaviors. We also introduce a mobility-specific end-to-end paradigm with variational attention for learning human mobility regularity, by which our GSCI can encode more meaningful and disentangled patterns into the trajectory representations. Besides, we design a graph neural network-based classifier to model the intrinsic associations among user connections which can significantly improve the inference performance. Extensive experiments conducted on real-world mobility datasets demonstrate the superiority of our proposed framework compared with the state-of-the-art approaches.