GC-MoE: Graph-Conditioned Mixture of Experts for Traffic Forecasting
The recently introduced framework, GC-MoE, allocates a tailored set of static forecasting specialists to each node, influenced by the graph's structure and current traffic data. This approach merges static pretrained spatio-temporal GNN specialists with an input-sensitive routing mechanism, focusing training solely on a lightweight routing component. Additionally, an optional layer for bounded graph-conditioned output refinement is explored. The effectiveness of this method is assessed across four benchmarks: PEMS04, PEMS07, METR-LA, and PEMS-BAY.
Key facts
- GC-MoE is a graph-conditioned mixture of experts framework.
- Each node gets a personalized combination of frozen forecasting experts.
- The combination is based on graph topology and recent traffic input.
- It uses frozen pretrained spatio-temporal GNN experts.
- An input-aware, spatially contextualized router is employed.
- Only a lightweight routing module is trained.
- A bounded graph-conditioned output refinement layer is optional.
- Evaluated on PEMS04, PEMS07, METR-LA, and PEMS-BAY.
Entities
—