Graph Ordering Attention Networks

Graph Neural Networks (GNNs) have been successfully used in many problems involving graph-structured data, achieving state-of-the-art performance. GNNs typically employ a message-passing scheme, which every node aggregates information from its neighbors using permutation-invariant aggregation function. Standard well-examined choices such as the mean or sum functions limited capabilities, they are not able to capture interactions among neighbors. In this work, we formalize these an information-theoretic framework that notably includes synergistic information. Driven by definition, introduce Ordering Attention (GOAT) layer, novel GNN component captures between nodes neighborhood. This is achieved learning local orderings via attention mechanism and processing ordered representations recurrent neural network aggregator. design allows us make use of permutation-sensitive aggregator while maintaining permutation-equivariance proposed GOAT layer. The model demonstrates increased performance modeling graph metrics complex information, betweenness centrality effective size node. practical use-cases, superior capability confirmed through success several real-world classification benchmarks.