Local High-order Structure-aware Graph Neural Network for motif prediction

Abstract

Motifs, serving as fundamental building blocks in complex networks, refer to small, frequently occurring connected subgraphs. Unlike link prediction, motif prediction focuses on whether a given set of nodes will form a particular type of motif and has achieved much more attention. Motif prediction holds significant research value and demonstrates broad application potential across various fields, such as financial default prediction and social network recommendation. However, existing research methods are relatively limited and have largely overlooked the crucial role of local higher-order correlations among nodes and the enclosing subgraph-level structural information in motifs. To overcome these challenges, we propose a novel Local High-order Structure-aware Graph Neural Network for motif prediction, named LHSGNN. It comprehensively predicts motifs from the perspectives of both the node level and subgraph level. LHSGNN incorporates local higher-order correlations among nodes to learn node embeddings. Then, it labels nodes from the views of node roles and distance metrics to capture complex structural information in enclosing subgraph-level motifs. Comprehensive experiments conducted on real-world datasets demonstrate the effectiveness of our LHSGNN.