DGNN: Decoupled Graph Neural Networks with Structural Consistency between Attribute and Graph Embedding Representations

TL;DR

This work tackles interference between node attributes and topology in graph neural networks by proposing DGNN, a decoupled framework that learns separate embeddings for attributes and graph structure while enforcing structural consistency via a shared reconstruction factor. By jointly modeling an attribute embedding, a topological graph embedding, and a semantic-graph embedding, DGNN constructs a comprehensive representation through concatenation, enabling improved node classification. An alternating-optimization scheme updates the multiple embeddings and reconstruction factors within an implicit-layer formulation, with complexity that scales quadratically in the number of nodes. Empirical results on six datasets show DGNN consistently outperforms strong baselines, especially on heterogeneous graphs, demonstrating the value of decoupling and structural-consistency regularization for robust graph representation learning.

Abstract

Graph neural networks (GNNs) demonstrate a robust capability for representation learning on graphs with complex structures, showcasing superior performance in various applications. The majority of existing GNNs employ a graph convolution operation by using both attribute and structure information through coupled learning. In essence, GNNs, from an optimization perspective, seek to learn a consensus and compromise embedding representation that balances attribute and graph information, selectively exploring and retaining valid information. To obtain a more comprehensive embedding representation of nodes, a novel GNNs framework, dubbed Decoupled Graph Neural Networks (DGNN), is introduced. DGNN explores distinctive embedding representations from the attribute and graph spaces by decoupled terms. Considering that semantic graph, constructed from attribute feature space, consists of different node connection information and provides enhancement for the topological graph, both topological and semantic graphs are combined for the embedding representation learning. Further, structural consistency among attribute embedding and graph embeddings is promoted to effectively remove redundant information and establish soft connection. This involves promoting factor sharing for adjacency reconstruction matrices, facilitating the exploration of a consensus and high-level correlation. Finally, a more powerful and complete representation is achieved through the concatenation of these embeddings. Experimental results conducted on several graph benchmark datasets verify its superiority in node classification task.

Figures (6)

• Figure 1: The overall framework of DGNN model, which simultaneously learns embedding representations from node attribute, topological graph, and semantic graph. The structural consistency aspect adeptly establishes the correlation between different representations, eliminating feature redundancy and achieving structural consistency.
• Figure 2: The experimental results of all ablation model and DGNN model on the all datasets, where A1: DGNN with $\lambda=0$ and $\varepsilon=0$, A2: DGNN with $\alpha=0$ and $\varepsilon = 1$, A3: DGNN with $\beta=0$.
• Figure 3: The t-SNE visualization for original attribute feature and node embedding representation obtained by DGNN in a 2-D space on the all datasets.
• Figure 4: The classification accuracy of DGNN with different $\varepsilon$ on the Cora, Citeseer, Chameleon and Squirrel datasets.
• Figure 5: The parameter sensitivity analysis of DGNN with respect to $\lambda$, $\alpha$, and $\beta$ on the Cora and Squirrel datasets.

Theorems & Definitions (1)

• Definition 1: Graph Signal Denoising (GSD)