Scholarlink

<p>Traditional graph neural networks&#xa0;(GNNs) typically understand graph data from a single view, which tends to over-look the multiple semantic information contained in graphs. However, the properties of graphs are usually determined by some substructures with varying sizes. Furthermore, stacking multiple convolutional layers in GNNs to capture high-order semantic information will lead to over-smoothing problem of graph embedding. Therefore, this paper proposes a multiple semantic-views and adaptive contrastive learning framework&#xa0;for graph classification, which could achieve better graph classification accuracy. In specific, our framework constructs multiple semantic views by incorporating node features that capture semantic information from neighborhoods of different sizes. Each view is processed with a customized number of convolutional layers. Then, we design an adaptive contrastive learning module to alleviate the over-smoothing problem. Finally, we use an attention mechanism to merge multiple semantic-views into the entire graph embedding. Comprehensive experiments show that the framework improves the classification performance and alleviates the problem of over-smoothing. Compared to the state-of-the-art methods, our approach achieves improvements of 1.7%, 2.5%, 8.4%, 3%, and 11.2% in accuracy on the NCI1, NCI109, PTC, ENZYMES, and PROTEINS datasets, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MVACL: joint multiple semantic-views and adaptive contrastive learning for graph classification

  • 期刊论文

  • 发表日期：

    2025年6月7日

查看全文
  • Zhigang Sun,
  • Jingjing Lei,
  • Li-e Wang,
  • Jinyong Sun,
  • Zhenkui Shi

摘要

Traditional graph neural networks (GNNs) typically understand graph data from a single view, which tends to over-look the multiple semantic information contained in graphs. However, the properties of graphs are usually determined by some substructures with varying sizes. Furthermore, stacking multiple convolutional layers in GNNs to capture high-order semantic information will lead to over-smoothing problem of graph embedding. Therefore, this paper proposes a multiple semantic-views and adaptive contrastive learning framework for graph classification, which could achieve better graph classification accuracy. In specific, our framework constructs multiple semantic views by incorporating node features that capture semantic information from neighborhoods of different sizes. Each view is processed with a customized number of convolutional layers. Then, we design an adaptive contrastive learning module to alleviate the over-smoothing problem. Finally, we use an attention mechanism to merge multiple semantic-views into the entire graph embedding. Comprehensive experiments show that the framework improves the classification performance and alleviates the problem of over-smoothing. Compared to the state-of-the-art methods, our approach achieves improvements of 1.7%, 2.5%, 8.4%, 3%, and 11.2% in accuracy on the NCI1, NCI109, PTC, ENZYMES, and PROTEINS datasets, respectively.