Multi-Modal Graph Interaction for Multi-Graph Convolution Network in Urban Spatiotemporal Forecasting

Xu Geng, Xiyu Wu, Lingyu Zhang, Qiang Yang, Yan Liu, Jieping Ye

Abstract

Urban spatiotemporal forecasting is critical for modern city management, yet it faces challenges in modeling complex, multi-modal spatial relationships and adapting to temporal distribution shifts. This paper introduces a two-tiered interaction mechanism for Multi-Graph Convolutional Networks (MGCN). By implementing Grouped GCNs (GGCN) for lower-level spatial features and Multi-linear Relationship GCNs (MRGCN) for high-level abstractions, the proposed framework enables effective "cross-talk" between different spatial modalities. Evaluation on real-world ride-hailing datasets demonstrates a 10% reduction in prediction error and significantly improved training efficiency and model robustness.

Introduction

Predicting urban dynamics, such as ride-hailing demand, requires understanding the non-Euclidean relationships between regions. Regions may be connected by physical proximity, similar Points of Interest (POIs), or infrastructure like road networks. While Multi-Graph models (MGCN) can process these modalities, they often treat them in isolation, failing to capture inter-modality interactions. Furthermore, urban patterns shift over time, leading to a "temporal shifting gap" where models overfit to training windows and fail to generalize as the test period drifts further from the training set.

Comparison of standard MGCN connectivity (a) versus the proposed compound graph connectivity (b) which allows for inter-modality interaction.

Methodology

The core contribution is a layered interaction strategy that distinguishes between physical connectivity and abstract statistical relationships across different depths of the neural network.

Grouped GCN (GGCN) for Lower Layers

To address the incompleteness of isolated graphs, GGCN facilitates information flow across modalities. It constructs a "compound graph" where nodes are connected not just within a single modality (e.g., road network) but across them (e.g., road network to POI similarity). Grouped sparsity regularization is applied to maintain model parsimony and prevent overfitting.

Multi-linear Relationship GCN (MRGCN) for Higher Layers

As features become more abstract in deeper layers, direct graph connections are replaced by statistical modeling. MRGCN treats modality weights as a tensor and imposes a tensor normal distribution prior. This encourages the model to learn coordinated, shared representations across all modalities, effectively targeting the temporal shifting gap.

The overall architecture of the proposed model, showing the transition from modality generation to layered GCN interactions and final aggregation.

Covariance Freezing

To enhance training stability and speed, the authors introduce covariance freezing for the input and output dimensions in the MRGCN. This technique prevents "co-adaptation" between neurons, leading to faster convergence and better generalization to unseen temporal distributions.

Experimental Evaluation

The model was tested using large-scale ride-hailing data from Didi Chuxing (Cities A and B) spanning March to December 2017. Auxiliary data included neighborhood grids, 13 POI categories, and road network data from OpenStreetMap.

Key Results

  • Predictive Accuracy: The combined GGCN and MRGCN approach achieved over a 10% reduction in Root Mean Squared Error (RMSE) compared to state-of-the-art baselines like STMGCN.
  • Efficiency: The model demonstrated superior data efficiency, reaching peak performance with only 3 months of training data compared to the 5 months required by baselines. Training speed also improved, with convergence occurring in 25-30 epochs versus 115 for standard models.
  • Robustness to Temporal Drift: By quantifying the temporal shifting gap using KL divergence, the authors showed that their multi-task learning approach maintains stable error rates even as test data moves 8 weeks away from the training period, whereas baseline errors spiked.

Conclusion

This research highlights that interaction in multi-modal graph networks should be depth-dependent. By combining physical compound graph connectivity in lower layers with statistical multi-linear relationships in higher layers, urban forecasting models can become more accurate, efficient, and robust to the evolving nature of city dynamics. Future work will explore the impact of graph density and the applicability of this framework to other urban tasks like air quality monitoring.

Summary updated

BibTeX
@inproceedings{geng2022multi,
  author = {Xu Geng and Xiyu Wu and Lingyu Zhang and Qiang Yang and Yan Liu and Jieping Ye},
  title = {{Multi-Modal Graph Interaction for Multi-Graph Convolution Network in Urban Spatiotemporal Forecasting}},
  booktitle = {Sustainability},
  year = {2022},
  doi = {10.3390/su141912397},
}
Figure