Enhancing Graph Neural Networks with Quantum Computed Encodings
Published in ICML 2024, 2023
We introduced novel positional encodings for graph transformers inspired by quantum systems, where long-range correlations emerge naturally from qubit interactions. By mapping graph topology to quantum circuit features, we developed expressive encodings that theoretically outperform standard methods like relative random walks for certain graph families. Empirically, we showed improved performance on benchmark datasets using tractable approximations of these quantum features. Best results were achieved on the chemical dataset ZINC.
Recommended citation: Slimane Thabet, Igor O. Sokolov, et al. (2023). "Enhancing Graph Neural Networks with Quantum Computed Encodings." ICML 2024
Download Paper