Learning node representation via Motif Coarsening

Motifs, as fundamental units of the graph, play a significant role in modeling complex systems in a variety of domains, including social networks, as well as biology and neuroscience. Motif preservation is a widely studied problem that provides new avenues for structure preservation. This paper is dedicated to exploring the significance of motifs with various patterns and effectively incorporating different motif patterns into node-level graph representation learning. We propose a novel node Representation learning framework via Motif Coarsening (RMC), which effectively incorporates different granularity structural information into node representation learning. RMC consists of two parallel components, the node representation learning aggregator and the motif-based node representation learning aggregator. In the node representation learning process, RMC directly encodes lower-order structures into node representation by a one-layer graph convolution network. For the motif-based node representation learning process, we propose a Motif Coarsening strategy for incorporating motif structure into the graph representation learning process. Furthermore, the MotifRe-Weighting strategy is proposed to biased convert motif representation into motif-based node representation. We verify the effectiveness of RMC by several node-related tasks on a series of widely used real-world datasets. Experimental results demonstrate that our proposed framework delivers superior promising representation performance to existing benchmarks. Ablation experiments proved that RMC has potential as an auxiliary framework, which indicates the excellent quality of Motif Coarsening and MotifRe-Weighting strategies over existing benchmarks from several evaluation metrics, involving mean classification accuracy, Micro-F1, and Macro-F1.