Daoud & El-Darieby's GNN layer study: ARMA and Chebyshev best for trajectory prediction

A new paper from researchers George Daoud and Mohamed El-Darieby, presented at the IEEE Intelligent Vehicles Symposium 2026, provides a comprehensive comparative study of 19 graph neural network (GNN) layer types for modeling spatiotemporal interactions in autonomous driving trajectory prediction. The study focuses on how different graph layers capture spatial interactions among road agents (e.g., vehicles, pedestrians) and temporal dynamics. Within their hyperparameter settings, five standout layer combinations emerged, with ARMA, Chebyshev, and topology-aware layers consistently delivering superior performance. The paper is notable for moving beyond raw metrics to extract practical design principles.

The key takeaways for engineers building trajectory prediction models: sum-based aggregation functions are significantly more effective than mean-based methods; multi-head attention mechanisms enable richer interaction modeling; and assigning different weights to different hop distances (i.e., how far away an agent is in the graph) markedly improves prediction accuracy. These findings help standardize the currently non-standardized GNN architecture design process for trajectory prediction. For autonomous driving teams, this translates to clearer guidelines on which layer types to prioritize (ARMA, Chebyshev, topology-aware) and which aggregation strategies (sum, multi-head attention, weighted hops) yield the best results, potentially leading to safer and more efficient self-driving systems.