<ul><li>InterGAT is proposed as a simplified alternative to Graph Attention Networks (GAT) for spatio-temporal forecasting, eliminating the need for predefined adjacency structures and dynamic attention scores.</li><li>InterGAT-GRU, incorporating a GRU-based temporal decoder, outperforms GAT-GRU in forecasting accuracy by 21% on the SZ-Taxi dataset and 6% on the Los-Loop dataset across all forecasting horizons.</li><li>The learned interaction matrix in InterGAT reveals interpretable, sparse, and topology-aware attention patterns aligning with community structure, enhancing understanding of functional topology driving predictions.</li><li>The model captures both localized and global dynamics, reducing training time by 60-70% compared to the GAT-GRU baseline, showcasing improved prediction accuracy, computational efficiency, and topological interpretability in dynamic graph-based domains.</li></ul>

Beyond Attention: Learning Spatio-Temporal Dynamics with Emergent Interpretable Topologies

Discover more