SADGCN-GC: self-attention-based deep graph convolutional neural network with quantization and visualization for graph classification
摘要
Since its establishment, Graph Neural Networks (GNNs) have shown great potential in processing graph-structured data. However, when increasing the depth of graph convolutional layers to improve classification performance, GNNs often suffer from over-smoothing, which leads to feature homogeneity and reduced model discrimination. To address this issue, we propose SADGCN-GC: a Self-Attention-based Deep Graph Convolutional Neural Network that integrates a self-attention pooling mechanism and DropEdge to effectively mitigate over-smoothing. Moreover, we introduce a novel metric called Prototypical Distance Ratio (PDR) to quantitatively measure the severity of over-smoothing and propose a T-SNE-based visualization module to enhance model interpretability. Experimental results on five benchmark datasets (DD, PROTEINS, NCI1, NCI109, REDDIT) demonstrate the superior performance of our model. For instance, SADGCN-GC improves classification accuracy by up to 6.72% on the DD dataset and reduces the PDR by 10.7% compared to existing self-attention pooling methods. These results validate the effectiveness and generalizability of SADGCN-GC in graph classification tasks. The code is available at https://github.com/ycitAi/SADGCN-GC.