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

TL;DR

REPLAY addresses location prediction over sparse human mobility trajectories by jointly modeling spatiotemporal context and time-varying temporal regularities. It introduces smoothed timestamp embeddings with timestamp-specific learnable bandwidths, integrated into a flashback-based RNN, and conditions predictions on a query timestamp. The approach yields 7.7–10.5% improvements over strong baselines and reveals interpretable patterns in temporal regularities, such as stronger morning regularities and weaker weekend patterns. This has practical impact for more accurate location recommendations and time-aware mobility services on sparse datasets.

Abstract

Location prediction forecasts a user's location based on historical user mobility traces. 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 feeding them as additional inputs to Recurrent Neural Networks (RNNs) or by using them to search for informative past hidden states for 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 periods, for example; this suggests the usefulness of the actual timestamps besides the temporal distances. Against this background, we propose REPLAY, a general RNN architecture learning to capture the time-varying temporal regularities for location prediction. Specifically, REPLAY not only resorts to the spatiotemporal distances in sparse trajectories to search for the informative past hidden states, but also accommodates the time-varying temporal regularities by incorporating 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. Our extensive evaluation compares REPLAY against a sizable collection of state-of-the-art techniques on two real-world datasets. Results show that REPLAY consistently and significantly outperforms state-of-the-art methods by 7.7\%-10.5\% in the location prediction task, and the bandwidths reveal interesting patterns of the time-varying temporal regularities.

Figures (4)

• Figure 1: Returning probability w.r.t. the temporal distances between check-ins. We plot the overall returning probability and its breakdown in the daytime (6:00-18:00) or nighttime (18:00-6:00). We observe that the daytime returning probability is significantly higher than in nighttime, which implies that the temporal regularity is much stronger in the daytime than in the nighttime.
• Figure 2: Overview of our proposed REPLAY. It consists of three components: 1) Smoothed timestamp embedding, 2) Flashback network with smoothed timestamp embeddings, and 3) Prediction conditioned on query timestamp.
• Figure 3: A toy example showing the key design ethos of REPLAY addressing the sparsity of trajectories by tackling three key challenges 1) varying and irregular time intervals, 2) uncertainty of the observed check-ins, and 3) unknown prediction time, through the three key components 1) Flashback mechanism, 2) smoothed timestamp embeddings, and 3) the query time, respectively.
• Figure 4: The learnt bandwidths and the corresponding location prediction performance across different timestamps.