GNN Layer Study Improves Autonomous Driving Trajectory Prediction.

Autonomous driving systems critically depend on accurate trajectory prediction for safe and efficient navigation. Graph Neural Networks (GNNs) are a promising tool for modeling the complex spatiotemporal interactions among various road agents. However, there's a lack of standardized guidance on which GNN layers are most effective for capturing these dynamics. This study conducts a comprehensive comparison of 19 different graph layer types, analyzing their capabilities in processing spatial interactions and temporal dynamics. The goal was to pinpoint the most effective architectural choices for trajectory prediction tasks. The research identified five standout layer combinations, with ARMA, Chebyshev, and topology-aware layers consistently outperforming others. Key design principles emerged, including the superiority of sum-based aggregation over mean-based methods, the benefit of multi-head attention for richer interactions, and the improved accuracy gained by assigning different weights to varying hop distances. These insights provide valuable guidance for developing more interpretable and effective trajectory prediction models.