Graph-Conditioned Mixture of Graph Neural Network Experts for Traffic Forecasting

Researchers propose GC-MoE, a graph-conditioned mixture of experts framework that improves traffic forecasting by assigning specialized neural network experts to different road segments based on graph topology. The approach trains only 17K parameters while leveraging 1.5M frozen expert weights, achieving competitive results across four standard traffic prediction benchmarks.

GC-MoE addresses a fundamental limitation in spatio-temporal forecasting: the assumption that a single model architecture can effectively handle diverse network dynamics. Road networks exhibit heterogeneous behavior—highways operate differently than residential streets, and congestion patterns vary by region. This research recognizes that node-wise specialization through expert assignment can capture these variations more efficiently than a monolithic approach.

The technical innovation lies in combining frozen pretrained GNN experts with a lightweight router that learns input-aware, spatially contextualized assignments. By training only the routing module (17K parameters) rather than full expert weights (1.5M parameters), the framework achieves parameter efficiency while maintaining competitive performance across PEMS04, PEMS07, METR-LA, and PEMS-BAY datasets. This design philosophy mirrors broader trends in machine learning toward modular, efficient architectures that leverage pretrained components.

For practitioners in urban computing and smart city infrastructure, this approach offers practical advantages: reduced computational overhead during training, flexible expert composition, and improved mean absolute error compared to ensemble baselines. The work demonstrates that specialized routing can outperform generic solutions without proportional increases in trainable parameters, suggesting scalability to larger networks.

The availability of open-source implementation enables rapid adoption and validation. Future developments likely include exploring dynamic expert selection based on real-time traffic conditions, extending the framework to multi-city scenarios, and investigating how graph structure changes affect routing decisions. This research reinforces the value of heterogeneity-aware modeling in complex networked systems.

• →GC-MoE assigns specialized neural network experts to road segments based on graph topology and traffic patterns, improving forecasting accuracy.
• →The framework trains only 17K parameters on top of 1.5M frozen expert weights, achieving significant parameter efficiency.
• →Achieves competitive or superior performance on four standard traffic prediction benchmarks (PEMS04, PEMS07, METR-LA, PEMS-BAY).
• →Graph-conditioned routing enables spatially contextualized expert selection rather than uniform model application across all nodes.
• →Open-source implementation facilitates broader adoption for urban computing and smart city traffic management applications.