Graph Attention MLP with Reliable Label Utilization

Graph neural networks (GNNs) have recently achieved state-of-the-art performance in many graph-based applications. Despite the high performance, they typically need to perform expensive recursive neighborhood expansions during every training epoch and has the scalability issue. Moreover, most of them are inflexible since they are restricted to fixed-hop neighborhoods and insensitive to actual receptive field requirements for each node. We circumvent these limitations by introducing a scalable and flexible method: Graph Attention Multi-Layer Perceptron (GAMLP). Following the routine of decoupled GNNs, the feature propagation in GAMLP is executed during pre-computation, which helps it maintain high scalability. With three proposed receptive field attention, each node in GAMLP is flexible in leveraging the propagated features over the different sizes of reception field. We conduct extensive evaluations on two large open graph benchmarks (ogbn-products and ogbn-papers100M), demonstrating that GAMLP not only achieves state-of-the-art performance, but also enjoys high scalability and efficiency.