SCARA

Recent advances in data processing have stimulated the demand for learning graphs of very large scales. Graph Neural Networks (GNNs), being an emerging and powerful approach solving graph tasks, are known to be difficult scale up. Most scalable models apply node-based techniques simplifying expensive message-passing propagation procedure GNN. However, we find such acceleration insufficient when applied million- or even billion-scale graphs. In this work, propose SCARA, a GNN with feature-oriented optimization computation. SCARA efficiently computes embedding from node features, further selects reuses feature computation results reduce overhead. Theoretical analysis indicates that our model achieves sub-linear time complexity guaranteed precision process as well training inference. We conduct extensive experiments on various datasets evaluate efficacy efficiency SCARA. Performance comparison baselines shows can reach up 100x than current state-of-the-art methods fast convergence comparable accuracy. notably, it is efficient precomputation largest available dataset Papers100M (111M nodes, 1.6B edges) 100 seconds.