High-Low Frequency Filtering Aware Graph Neural Networks with Representation Decoupling Learning

Graph neural networks (GNNs) perform excellently in representation learning of graph data, attracting widespread attentions. Although, most GNNs with explicit graph convolution face the interference between attribute and topological structure, resulting in partial and compromise embedding representation. Most GNNs only capture the low-frequency information of graph signals and ignore the important high-frequency information. To address these issues, we propose a novel framework for GNNs, i.e., High-Low Frequency Filtering aware Graph Neural Network (HLFF-GNN). Under the designed optimization framework, HLFF-GNN learns various representations to fit attribute and graph structure with representation decoupling learning strategy, which obtains more comprehensive representation by concatenating them. High and low frequency information is simultaneously exploited, where especially the high-frequency information would be additional and useful supplement to low-frequency information. Further, diversity constraint is introduced to enhance complementarity of various embedding representations. The experimental results of node classification task on several benchmark datasets demonstrate the validity and superiority of our approach.