AFMF: adaptive fusion of multi-hop neighborhood features in graph convolutional network

Abstract

Abstract Graph-structured data has been widely used in modern information management systems. Effectively extracting the latent structural and semantic relationships between nodes in the graph is a key research challenge. Graph Neural Networks (GNNs) can efficiently process homophilic graphs by learning node embedding with neighborhood information. However, heterophilic graphs where adjacent nodes may belong to different categories are ubiquitous and cannot be directly processed by GNNs. To address this issue, we propose a graph convolutional network based on Adaptive Fusion of Multi-hop Features, termed AFMF, which can adaptively generate feature fusion weights according to the features of multi-hop nodes, solving the over-smoothing problem caused by shared fusion weights. It adopts two different message passing strategies to aggregate node features of multi-hop neighbors, based on which weights are adaptively generated to fuse the aggregated node features for node classification. More specifically, we propose a multi-hop neighborhood feature extraction module (MNFE), which constructs multi-hop adjacency matrices with and without self-loops for extracting node features. Then, we propose a feature fusion weight generation module (FFWG), which can adaptively generate fusion weights based on the above aggregated node features. Finally, the feature fusion module (FF) applies the adaptively generated weights to fuse the multi-hop neighborhood features of each node to get the final node representation. The cross entropy loss function drives the network to preserve critical information. Extensive experiments over eight benchmark datasets with different heterophilic levels demonstrate that our proposed method can offer comparable or superior state-of-the-art performance on both homophilic and heterophilic graphs.

Keywords

• Graph neural network
• Multi-hop neighborhood features
• Adaptive feature fusion
• Heterophilic graphs
• Node classification