SLAPS: Self-Supervision Improves Structure Learning for Graph Neural Networks

Graph neural networks (GNNs) work well when the graph structure is provided.\nHowever, this structure may not always be available in real-world applications.\nOne solution to this problem is to infer a task-specific latent structure and\nthen apply a GNN to the inferred graph. Unfortunately, the space of possible\ngraph structures grows super-exponentially with the number of nodes and so the\ntask-specific supervision may be insufficient for learning both the structure\nand the GNN parameters. In this work, we propose the Simultaneous Learning of\nAdjacency and GNN Parameters with Self-supervision, or SLAPS, a method that\nprovides more supervision for inferring a graph structure through\nself-supervision. A comprehensive experimental study demonstrates that SLAPS\nscales to large graphs with hundreds of thousands of nodes and outperforms\nseveral models that have been proposed to learn a task-specific graph structure\non established benchmarks.\n