Disentangling and Re-evaluating Similarity-Based Graph Structure Learning for GNNs on Node Classification

Abstract: Graph Structure Learning (GSL) is widely used to improve Graph Neural Networks (GNNs), especially through similarity-based graph construction for node classification. However, it remains unclear whether the reported gains come from the learned graph itself or from the node representations used to build that graph. In this paper, we study this question through a framework that decomposes GSL into three steps: (1) GSL-base generation (\ie processed node embeddings), (2) graph construction, and (3) multi-view fusion. Through empirical analysis and theoretical results, we show that, in the similarity-based setting, graph convolution on the constructed graph does not increase the Mutual Information (MI) between node representations and labels. This suggests that improvements often come from the quality of the GSL bases rather than from the graph construction procedure. To test this claim, we evaluate 450 GSL variants and compare them with GNN baselines under a shared search space of GSL bases. In this setting, similarity-based graph construction provides limited or inconsistent gains, whereas strong pre-trained GSL bases account for most of the improvement. These results clarify which components of GSL matter most for node classification and suggest that simpler GSL designs may be sufficient in many cases.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Moshe_Eliasof1

Submission Number: 8087