Disentangling and Re-evaluating The Effectiveness of Graph Structure Learning For GNNs

Yilun Zheng, Zhuofan Zhang, Ziming Wang, Xiang Li, Sitao Luan, Xiaojiang Peng, Lihui Chen

Keywords: Graph Neural Networks, Graph Representation Learning, Graph Structure Learning

Abstract: Graph Structure Learning (GSL) has been widely adopted in the design of Graph Neural Networks (GNNs), with similarity-based graph learning emerging as the most popular approach for node classification. However, which component of GSL really enhances GNN performance remains underexplored. In this paper, we disentangle its effects and present a comprehensive analysis. Specifically, we propose a novel framework that can decompose GSL into three steps: (1) GSL bases (i.e. processed node embeddings for construction) generation, (2) new graph construction, and (3) view fusion. Through empirical studies and theoretical analysis, we demonstrate that applying graph convolution to the newly constructed graphs does not increase the Mutual Information (MI) between node embeddings and labels. Our findings reveal that model performance is primarily driven by the quality of GSL bases rather than the graph construction methods. To validate them, we conduct extensive experiments with 450 GSL variants and benchmark them against GNN baselines within the same search space for GSL bases. Results show that similarity-based graph construction has negligible or even adverse impacts on GNN performance, while pre-trained GSL bases provide significant performance gains. These findings verify and confirm our analysis, underscoring the critical role of GSL bases and highlighting the need to simplify the other two GSL steps.

Primary Area: Evaluation (e.g., data collection methodology, data processing methodology, data analysis methodology, meta studies on data sources, extracting signals from data, replicability of data collection and data analysis and validity of metrics, validity of data collection experiments, human-in-the-loop for data collection, human-in-the-loop for data evaluation)

Submission Number: 1680