Topological-Semantic Structure and Attribute Representations Decoupling Learning for Improved Graph Neural Networks

Graph neural networks (GNNs) have demonstrated strong embedding representation learning capabilities on graph with complex structure and have achieved significant success in many applications. The GNNs with explicit graph convolution usually learn a unified representation encoding both attribute feature and graph structure. However, attribute and graph structure are different characteristics of nodes. The unified representation is a compromised representation, which only contains partial information shared by attribute and graph. In addition, most current GNNs also suffer from the problem of ignoring semantic topology. In this paper, to alleviate the interference and redundancy of representation learning, as well as the problem of insufficient structural information, we propose a new Representation Decoupling Learning for Improved Graph Neural Networks (IGNN-RDL) with optimization framework, which learns various representations for attribute and topological-semantic graphs. Further, a diversity constraint is introduced to enhance the complementarity of various representations, resulting in a comprehensive embedding representation by concatenating them. IGNN-RDL iteratively updates attribute representation and graph embedding representation, which can avoid multilayer stacking of graph convolution and alleviate over-smoothing issue. Experimental results on multiple graph benchmark datasets validate the superiority of the proposed IGNN-RDL in node classification tasks.