A Temporal-Context-Aware Approach for Individual Human Mobility Inference Based on Sparse Trajectory Data

Abstract

Inferring individual human mobility at a given time is not only beneficial for personalized location-based services, but also crucial for trajectory tracking of the confirmed cases in the context of the COVID-19 pandemic. However, individual generated trajectory data using mobile Apps is characterized by implicit feedback, which means only a few individual-location interactions can be observed. Existing studies based on such sparse trajectory data are not sufficient to infer individual’s missing mobility in his/her historical trajectory and further predict individual’s future mobility given a specific time. To address this concern, in this paper, we propose a temporal-context-aware approach that incorporates multiple factors to model the time sensitive individual-location interactions in a bottom-up way. Based on the idea of feature fusion, the driving effect of heterogeneous information such as time, space, category and sentiment on individual’s mobile behavior is gradually strengthened, so that the temporal context when a check-in occurs can be accurately depicted. We leverage Bayesian Personalized Ranking (BPR) to optimize the model, where a novel negative sampling method is employed to alleviate data sparseness. Based on three real-world datasets, we evaluate the proposed approach with regard to two different tasks, namely, missing mobility inference and future mobility prediction at a given time. The empirical results encouragingly demonstrate that our approach outperforms multiple baselines in terms of two evaluation metrics, i.e., accuracy and average percentile rank.