TL;DR
This paper introduces SCOT, a novel framework for transferring knowledge between cities to improve predictions in data-scarce urban areas. SCOT uses optimal transport to learn explicit, soft correspondences between regions in different cities, overcoming limitations of previous methods and achieving significant accuracy improvements across various urban computing tasks.
What Problem Does This Paper Address?
Urban computing tasks, like predicting regional GDP or population, often rely on data from well-instrumented cities to make predictions in cities with limited data. This "cross-city transfer" is challenging because cities rarely have perfectly aligned regions (e.g., different numbers of administrative districts) and the underlying patterns (like commuting routes) might not perfectly transfer. Existing methods struggle with this:
- Anchor Matching: Relies on manually defined or heuristically matched regions, which are brittle when region counts differ or when some regions become "hubs" that absorb too much information.
- Distribution Matching: Aligns entire city-wide embedding distributions, which can over-mix embeddings and obscure city-specific nuances.
The core problem SCOT addresses is the lack of a robust mechanism to learn explicit, yet flexible, correspondences between unequal region sets across cities, especially when dealing with multiple source cities.
Methods
SCOT introduces a novel alignment mechanism based on Optimal Transport (OT), specifically using the Sinkhorn algorithm for efficient, differentiable computation. Here's a breakdown of its key components:
- Sinkhorn-Based Soft Correspondence: Instead of hard one-to-one matches, SCOT learns a "soft" correspondence matrix (P) that assigns probabilities of association between regions in different cities. This is achieved by minimizing an OT cost matrix (based on Euclidean distance between region embeddings) with entropic regularization. This prevents the "hubness" problem by distributing mass and avoids over-mixing by maintaining explicit region-level connections.
- OT-Guided Contrastive Objective: To ensure the learned correspondences are semantically meaningful, SCOT uses the OT coupling matrix (P) to weight a contrastive loss. This encourages embeddings of corresponding regions to be similar while pushing dissimilar regions apart, sharpening the transferable structure.
- Cycle Reconstruction Regularizer: To stabilize training, especially under significant city heterogeneity, SCOT employs a one-sided cycle reconstruction. This ensures that a region's semantic representation can be recovered after passing through the target city's representation and back, reinforcing explicit correspondences without assuming symmetry.
- Multi-Source Hub Alignment: For scenarios with multiple source cities, SCOT introduces a shared prototype hub. Instead of aligning each source to the target independently, all cities (sources and target) are aligned to this hub. This hub acts as a common latent space, mediated by a target-induced prototype prior that guides prototypes towards target-relevant semantics. This approach prevents source domination and mitigates conflicting gradients from multiple sources.
Datasets
The paper uses real-world urban mobility data from three Chinese cities:
- Beijing (BJ)
- Xi'an (XA)
- Chengdu (CD)
This data is used to form directed mobility graphs and is evaluated on three region-level prediction tasks:
- Gross Domestic Product (GDP)
- Population
- Carbon Emissions (CO2)
The evaluation also includes cross-country transfer experiments with New York City (NYC), highlighting the framework's generalization capabilities.
Key Insights and Ideas
- Explicit Soft Correspondences are Key: The paper strongly argues that the alignment mechanism, not just the GNN encoder's capacity, is the bottleneck for cross-city transfer. SCOT's success hinges on learning explicit, yet flexible, soft correspondences.
- Optimal Transport as a Primitive: OT, particularly with entropic regularization, provides a principled way to handle unequal partitions and learn mass-controlled alignments, addressing the limitations of prior methods.
- Contrastive Sharpening for Semantics: Geometric alignment alone isn't enough. The OT-guided contrastive loss is crucial for making the aligned representations semantically discriminative and transferable.
- Hub for Multi-Source Harmony: A shared prototype hub is a powerful way to aggregate knowledge from multiple, potentially conflicting, sources. The target-induced prior is essential for steering this hub effectively and preventing source domination.
- Robustness to Hyperparameters: SCOT demonstrates stable performance across a wide range of hyperparameters, making it practical for real-world applications where extensive tuning is infeasible.
Key Results
- Significant Accuracy Gains: SCOT consistently outperforms strong baselines in both single-source (5-50% relative MAE/MAPE reduction) and multi-source settings.
- Multi-Source Advantage: Multi-source SCOT consistently surpasses the strongest single-source transfer, demonstrating the value of aggregating diverse urban data.
- Encoder Agnostic: The gains are primarily from the alignment design, not the choice of GNN encoder, as evidenced by ablation studies showing much larger performance differences due to alignment changes than encoder substitutions.
- Cross-Country Generalization: SCOT successfully transfers knowledge to and from New York City, a city with significantly different urban form and mobility patterns, showcasing its robustness.
- Interpretable Diagnostics: The learned OT couplings and hub assignments provide interpretable insights into the alignment process and potential sources of conflict.
Limitations and Future Work
- Interpretability vs. Causality: While SCOT provides interpretable diagnostics (e.g., conflict signals), these are not guarantees of causal correctness and should be used with domain knowledge.
- Scope of Experiments: The current experiments focus on mobility-derived graphs and aggregated socioeconomic targets. Extending the framework to finer spatial resolutions or different urban modalities might require additional modeling assumptions.
- Future Work: The authors suggest exploring uncertainty-aware mechanisms for selectively integrating sources under severe heterogeneity and developing more sophisticated conflict-aware aggregation strategies for multi-source transfer.
Summary updated