Gate the Filter, Not the Message: Node-Channel Mixtures for Pre-Propagation GNNs
Researchers have developed a new approach for pre-propagation graph neural networks (PPGNNs) called FilterMoE. This method addresses the puzzle of why more complex aggregators don't always outperform simpler ones in PPGNNs. FilterMoE introduces a mixture-of-experts design that routes Chebyshev filter experts jointly over nodes and channels, outperforming existing PPGNNs on nine out of eleven benchmarks. AI
IMPACT Introduces a novel routing mechanism for graph neural networks, potentially improving performance on various graph-based tasks.