Capturing High-order Structures on Motif-based Graph Nerual Networks

TL;DR

This paper introduces a motif-based graph neural network framework that captures high-order structures to improve node representation, addressing limitations of shallow and deep embeddings in traditional GNNs.

Contribution

The proposed method leverages network motifs for deep feature learning, enhancing the ability to represent high-order network structures in GNNs.

Findings

Significant improvement in link prediction accuracy

Enhanced node classification performance

Effective capturing of high-order network structures

Abstract

Graph Nerual Networks (GNNs) are effective models in graph embedding. It extracts shallow features and neighborhood information by aggregating neighbor information to learn the embedding representation of different nodes. However, the local topology information of many nodes in the network is similar, the network obtained by shallow embedding represents the network that is susceptible to structural noise, and the low-order embedding cannot capture the high-order network structure; on the other hand, the deep embedding undergoes multi-layer convolution. After the filters are stacked, the embedded distribution is destroyed, and graph smoothing occurs. To address these challenges, we propose a new framework that leverages network motifs to learn deep features of the network from low-level embeddings under the assumption of network homogeneity and transitivity, and then combines local neighborhood information with deeper global information fusion results in accurate representation of nodes. On real datasets, our model achieves significant performance improvement in link prediction and node classification.