Port-Hamiltonian Architectural Bias for Long-Range Propagation in Deep Graph Networks
Published in 2025 International Conference on Learning Representations (ICLR), 2025
This work tackles the challenge of long-range information propagation in graph neural networks. We proposed port-Hamiltonian Deep Graph Networks, a novel framework inspired by physics that models neural information flow using principles from Hamiltonian systems. By combining conservative and dissipative dynamics, we control how information spreads or fades across the graph. Despite its theoretical grounding, the method integrates easily into standard message-passing networks and shows strong empirical performance, especially in tasks that require effective long-range reasoning. Download paper here