Abstract
Understanding urban mobility through Origin-Destination (OD) crowd flow is critical for effective city planning. However, real-world OD data often suffers from extreme sparsity, noise, and a lack of interpretability in deep learning models. This paper introduces UApex, a framework that combines a Variational Multi-modal Recurrent Graph Auto-Encoder (VMR-GAE) with a novel explainability module to provide robust and transparent OD flow interpolation.
The Challenge of OD Flow Interpolation
Origin-Destination (OD) crowd flow data tracks the movement of people between specific regions in a city. Managing this data presents three primary challenges:
- Sparsity: At any given time, the majority of possible OD pairs have zero or missing values.
- Stochasticity: Human mobility is inherently uncertain; events, holidays, and weather can cause significant deviations from historical norms.
- Interpretability: High-performing deep learning models often act as "black boxes," making it difficult for urban planners to trust their outputs for critical infrastructure decisions.
The UApex Framework
To address these issues, the authors propose the UApex framework, which operates in two distinct stages: interpolation and explanation.
1. The Interpolator: VMR-GAE
The core of UApex is the Variational Multi-modal Recurrent Graph Auto-Encoder (VMR-GAE). Unlike standard models that predict deterministic values, VMR-GAE learns a probability distribution of latent variables, allowing it to quantify uncertainty.
Key technical components include:
- Diffusion Convolution & Recurrence: These layers capture the directed nature of OD flow (where movement from A to B is distinct from B to A) and its temporal dependencies.
- Multi-modal Distribution Alignment: The model integrates "context modals" (e.g., supplementary taxi data) to fill data gaps, using an alignment mechanism to fuse disparate data sources in the latent space.
- Poisson Prior Decoder: Recognizing that crowd flow consists of discrete counts, the authors utilize a Poisson distribution in the decoder, which offers superior mathematical alignment with the data compared to standard Gaussian distributions.
2. The Explainer
To ensure transparency, UApex incorporates an explainability module based on Shapley values from cooperative game theory. This module identifies which historical states or subgraphs contributed most to a specific interpolation. To make this computationally feasible for large-scale urban graphs, the authors developed an optimized exploration algorithm using Monte Carlo sampling and parallel processing.
Key Insights and Results
The UApex framework was evaluated on real-world datasets from Beijing and New York City. The results highlighted several key advantages:
- Superior Accuracy: VMR-GAE consistently outperformed baselines such as 2D-GCN and DeepEnsembles. The improvements were most pronounced during high-volatility periods like rush hours and holidays, where it improved Mean Absolute Error (MAE) by over 17% on the Beijing dataset.
- Directed Graph Dynamics: By treating regions as nodes in a directed graph, the model successfully captured the inherent asymmetry of urban flow—for instance, how morning commute patterns differ fundamentally from evening returns.
- The "Living Circle" Effect: The explainer revealed that the model identifies "living circles"—clusters of residential and commercial areas that historically influence each other. This confirms that the model learns actual urban structures rather than merely memorizing statistical noise.
Conclusion
UApex represents a significant step forward in urban computing by treating uncertainty as a feature rather than a nuisance. By combining robust variational inference with game-theoretic explainability, the framework provides city planners with a tool that is not only accurate but also transparent and reliable under the unpredictable conditions of real-world urban mobility.
Summary updated