interaction-modelling

This paper compares 19 graph neural network layer types for modelling interactions in driving trajectory prediction, finding ARMA, Chebyshev, and topology-aware layers most effective and offering design principles for better prediction models.