MoGERNN: An Inductive Traffic Predictor for Unobserved Locations

Abstract

Given a partially observed road network, how can we predict the traffic state of interested unobserved locations? Traffic prediction is crucial for advanced traffic management systems, with deep learning approaches showing exceptional performance. However, most existing approaches assume sensors are deployed at all locations of interest, which is impractical due to financial constraints. Furthermore, these methods are typically fragile to structural changes in sensing networks, which require costly retraining even for minor changes in sensor configuration. To address these challenges, we propose MoGERNN, an inductive spatio-temporal graph model with two key components: (i) a Mixture of Graph Experts (MoGE) with sparse gating mechanisms that dynamically route nodes to specialized graph aggregators, capturing heterogeneous spatial dependencies efficiently; (ii) a graph encoder-decoder architecture that leverages these embeddings to capture both spatial and temporal dependencies for comprehensive traffic state prediction. Experiments on two real-world datasets show MoGERNN consistently outperforms baseline methods for both observed and unobserved locations. MoGERNN can accurately predict congestion evolution even in areas without sensors, offering valuable information for traffic management. Moreover, MoGERNN is adaptable to the changes of sensor network, maintaining competitive performance even compared to its retrained counterpart. Tests performed with different numbers of available sensors confirm its consistent superiority, and ablation studies validate the effectiveness of its key modules. The code of this work is publicly available at: https://github.com/ZJU-TSELab/MoGERNN.

References in corpus (16)

• Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting
• T-GCN: A Temporal Graph ConvolutionalNetwork for Traffic Prediction
• Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks
• Deep Learning on Traffic Prediction: Methods, Analysis and Future Directions
• Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting
• TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis
• Traffic Prediction using Artificial Intelligence: Review of Recent Advances and Emerging Opportunities
• SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction
• Principal Neighbourhood Aggregation for Graph Nets
• A Physics-Informed Deep Learning Paradigm for Traffic State and Fundamental Diagram Estimation
• Macroscopic Traffic Flow Modeling with Physics Regularized Gaussian Process: A New Insight into Machine Learning Applications
• Kriging Convolutional Networks
• Correlating sparse sensing for large-scale traffic speed estimation: A Laplacian-enhanced low-rank tensor kriging approach
• Spatial Aggregation and Temporal Convolution Networks for Real-time Kriging
• Understanding and Mitigating Extrapolation Failures in Physics-Informed Neural Networks
• KITS: Inductive Spatio-Temporal Kriging with Increment Training Strategy