TL;DR
This paper introduces MSTDP, a novel model for predicting human mobility patterns for the next day (mid-term prediction), going beyond traditional next-location forecasting. By decoupling location and duration, using hierarchical temporal encoders, and a heterogeneous spatial graph, MSTDP significantly improves accuracy in urban planning and epidemic modeling applications.
What Problem Does This Paper Address?
Current human mobility prediction models primarily focus on predicting the very next location an individual will visit. While useful for personalized services like recommendations, this "next-location prediction" is insufficient for broader societal applications such as urban planning, traffic management, and epidemic control. These applications require forecasting an individual's entire daily travel trajectory for the upcoming day or even week. This "mid-term mobility prediction" is challenging due to:
- Location Repetition: Individuals often stay at the same location for extended periods (e.g., home overnight), leading to redundant data points in trajectories.
- Long Sequences: Daily trajectories can be very long, making it difficult for models to capture long-term dependencies and patterns.
- Multi-scale Spatiotemporal Patterns: Human mobility exhibits both daily recurrence (e.g., commuting patterns) and weekly periodicity, as well as complex spatial relationships between urban areas at different scales. Existing models struggle to effectively capture these multi-faceted patterns.
Methods
The paper proposes the Multi-scale Spatial-Temporal Decoupled Predictor (MSTDP), which tackles these challenges through several key innovations:
- Spatial-Temporal Decoupler: This module addresses location repetition by separating a daily trajectory into two distinct chains: a location chain (unique locations visited) and a duration chain (time spent at each location). This effectively "decouples" the spatial and temporal aspects and shortens sequences.
- Hierarchical Trajectory Encoder: To capture multi-scale temporal patterns, a two-layer encoder is used. A daily encoder processes daily travel transitions, while a weekly encoder captures weekly variations, allowing the model to learn from patterns observed a week prior.
- Spatial Heterogeneous Graph Learner (Geospatial Embedder): This component models complex spatial relationships. It constructs a heterogeneous graph that integrates grid-level and administrative-level data, considering both flow (movement between areas) and adjacency (proximity) relationships. This provides richer, multi-scale semantic representations of locations.
- Transformer-based Decoders: Both spatial and temporal decoders utilize Transformer architectures to effectively attend to historical information and make iterative predictions for the location and duration chains, respectively.
Datasets
The study utilizes five real-world urban-scale datasets:
- Boston
- Los Angeles (LA)
- San Francisco Bay Area (SFBay)
- Shanghai
- Tokyo
These datasets are derived from mobile phone records (Call Detail Records - CDRs). The paper notes that the Boston, LA, and SFBay datasets use census tracts as administrative units, Shanghai uses towns, and Tokyo uses grid squares due to blurred coordinates. For the Tokyo dataset, specific preprocessing was done to identify home locations and filter users with fewer than 10 unique locations per day.
Key Insights and Ideas
- Mid-term prediction is a distinct and important problem: The paper clearly articulates why forecasting a full day's trajectory is crucial for societal applications and how it differs from next-location prediction.
- Decoupling location and duration is key: By separating "where" from "how long," the model can better learn underlying travel patterns without being confounded by repeated location visits.
- Hierarchy matters for temporal patterns: Capturing both daily and weekly regularities is essential for robust mid-term prediction, and a hierarchical encoder effectively achieves this.
- Spatial relationships are multi-layered: Integrating grid-level and administrative-level spatial information in a heterogeneous graph provides a more comprehensive understanding of urban structure and mobility.
- Model interpretability through statistical physics: The use of statistical physics analysis (motif analysis, travel distance distribution) provides deeper insights into why the model performs well, showing it captures real-world mobility behaviors better than baselines.
Key Results
- Superior Performance: MSTDP consistently outperforms 11 baseline models across various metrics (Accuracy, DevDist, TravelDist, DepartTime) on five real-world datasets. For example, on the Boston dataset, MSTDP achieved a 3.6% improvement in Accuracy and a 9.0% reduction in DevDist compared to the best baseline.
- Epidemic Modeling Impact: When applied to epidemic transmission modeling in Boston, MSTDP significantly reduced the Mean Absolute Error (MAE) for predicting active cases by 62.8% and cumulative new cases by 87.6% compared to the best baseline (HTAED).
- Robustness to Forecasting Horizon: MSTDP shows slower performance degradation than baselines when forecasting further into the future (up to 7 days), attributed to its broader contextual awareness.
- Efficient Training and Inference: While not the absolute fastest, MSTDP demonstrates competitive efficiency, significantly reducing training time compared to other advanced models like MTNet and AGRAN.
Limitations and Future Work
- Generalization to Sparse Data: MSTDP performs strongly on regular movement patterns and dense datasets. Its generalization to irregular movement scenarios or sparse data environments is an area for future exploration.
- Privacy Concerns: The use of mobile phone data raises privacy concerns. Future work could explore advanced privacy-preserving mechanisms.
- External Factors: Real-world mobility is influenced by external factors like weather, public health events, and holidays. Integrating these and developing dynamic adaptation mechanisms could improve robustness.
- Cumulative Error Propagation: Mid-term prediction inherently faces challenges with cumulative error over multiple steps. Future research could focus on designing methodologies to further mitigate this.
Summary updated