REPLAY: Modeling Time-Varying Temporal Regularities of Human Mobility for Location Prediction Over Sparse Trajectories

Next-location prediction aims to forecast which location a user is most likely to visit given the user’s historical data. As a sequence modeling problem by nature, it has been widely addressed using Recurrent Neural Networks (RNNs). To tackle the intrinsic sparsity issue of real-world user mobility traces, spatiotemporal contexts have been shown as significantly useful. Existing solutions mostly incorporate spatiotemporal distances between locations in mobility traces, either by integrating them into the RNN units as additional information, or utilizing them to search for informative historical hidden states to improve prediction. However, such distance-based methods fail to capture the time-varying temporal regularities of human mobility, where human mobility is often more regular in the morning than in other time periods, for example; this suggests the usefulness of the actual timestamps besides the temporal distances. Under this circumstance, we propose REPLAY, learning to capture the time-varying temporal regularities for location prediction based on general RNN architecture. Specifically, REPLAY is designed on top of a flashback mechanism, where the spatiotemporal distances in sparse trajectories are used to search for the informative past hidden states; to accommodate the time-varying temporal regularities, REPLAY incorporates smoothed timestamp embeddings using Gaussian weighted averaging with timestamp-specific learnable bandwidths, which can flexibly adapt to the temporal regularities of different strengths across different timestamps. We conduct a comprehensive evaluation, comparing REPLAY against a wide range of state-of-the-art methods. Experimental results show REPLAY significantly and consistently outperforms state-of-the-art methods by 7.7%–10.5% in the location prediction task, and the learnt bandwidths reveal interesting patterns of the time-varying temporal regularities.