Saliency-Aware Regularized Graph Neural Network

TL;DR

The paper addresses graph classification by tackling two challenges: explicit modeling of global node saliency and improving graph-level representations. It introduces SAR-GNN, which combines a backbone GNN with a Graph Neural Memory that distills a compact graph representation and guides saliency-aware regularization of neighborhood aggregation. Through iterative, end-to-end learning, the memory and backbone mutually refine each other, yielding more informative graph representations and interpretable node saliency maps. Across seven datasets and multiple backbones, SAR-GNN delivers consistent gains over baselines and competitive state-of-the-art performance, demonstrating the utility of explicit saliency modeling for graph classification.

Abstract

The crux of graph classification lies in the effective representation learning for the entire graph. Typical graph neural networks focus on modeling the local dependencies when aggregating features of neighboring nodes, and obtain the representation for the entire graph by aggregating node features. Such methods have two potential limitations: 1) the global node saliency w.r.t. graph classification is not explicitly modeled, which is crucial since different nodes may have different semantic relevance to graph classification; 2) the graph representation directly aggregated from node features may have limited effectiveness to reflect graph-level information. In this work, we propose the Saliency-Aware Regularized Graph Neural Network (SAR-GNN) for graph classification, which consists of two core modules: 1) a traditional graph neural network serving as the backbone for learning node features and 2) the Graph Neural Memory designed to distill a compact graph representation from node features of the backbone. We first estimate the global node saliency by measuring the semantic similarity between the compact graph representation and node features. Then the learned saliency distribution is leveraged to regularize the neighborhood aggregation of the backbone, which facilitates the message passing of features for salient nodes and suppresses the less relevant nodes. Thus, our model can learn more effective graph representation. We demonstrate the merits of SAR-GNN by extensive experiments on seven datasets across various types of graph data. Code will be released.

Figures (8)

• Figure 1: Architecture of the Saliency-Aware Regularized Graph Neural Network (SAR-GNN). It consists of two core modules: 1) a traditional graph neural network serving as the backbone network for learning node features and 2) the Graph Neural Memory for distilling a compact graph representation from node features of the backbone. The learned graph representation is leveraged to measure the global node saliency and regularize the backbone. Thus, two modules work interdependently to refine node features and the compact graph representation in an iterative manner. The operation of Saliency-Aware Regularization in the left panel is elaborated in the right panel.
• Figure 2: The performance of our SAR-GNN adopting two optimization policies, namely the joint training policy and the alternating training policy, on MUTAG (a) and Letter-high datasets (b), respectively.
• Figure 3: Visualization of two randomly selected subgraphs generated by DSS-GNN (EGO+) on samples from MUTAG and TRIANGLES datasets, respectively. The triangles are indicated in shaded regions for the TRIANGLES sample. While DSS-GNN can preserve the triangle structure in the generated subgraphs, its generated subgraphs on the MUTAG sample fail to preserve intact fused rings whose number is closely correlated to the recognition of mutagenic aromatic.
• Figure 4: Visualization of learned node saliency by our SAR-GCN, ChebyGIN-unsup and SAGPool on three graphs randomly selected from TRIANGLES test data. The triangles are indicated in shaded regions and the number of triangles is given as the graph label. Our SAR-GCN tends to assign higher saliency scores to those nodes associated with more triangles, and thus shows higher accuracy than ChebyGIN-unsup and SAGPool in capturing the discriminative nodes associated with more triangles. The nodes that are assigned distinctly inaccurate saliency weights are marked with red dotted circles.
• Figure 5: Visualization of learned node saliency by our SAR-GraphSAGE, ChebyGIN-unsup and SAGPool on three graphs, labeled as 'mutagenic aromatic', randomly selected from MUTAG. Our SAR-GraphSAGE can capture the key nodes located at the center of fused rings which are discriminative for 'mutagenic aromatic', whilst ChebyGIN-unsup and SAGPool either assign roughly uniform attention to all nodes or fail to capture the key nodes.