Scholarlink

<p>Graph representation learning has emerged as a powerful approach for modeling structured data across diverse domains, including social networks, biochemical interactions, and financial transaction systems. Existing contrastive learning frameworks have demonstrated strong performance in self-supervised graph learning; however, they often suffer from suboptimal augmentations, inadequate structural preservation, and a lack of adaptability to heterogeneous graph structures such as blockchain transaction networks. In this work, we propose <i>ACLGMI</i> (Adaptive Contrastive Learning with Graph Mutual Information Maximization), a novel framework that extends traditional contrastive learning by introducing an adaptive augmentation strategy and a mutual information maximization objective to enhance representation quality. ACLGMI dynamically adjusts augmentation strategies based on graph topology, ensuring that critical structural properties are preserved while improving robustness to noise and adversarial perturbations. Furthermore, we introduce a multi-level mutual information constraint that maximizes global and local consistency in learned representations, leading to improved performance across multiple downstream tasks. We evaluate ACLGMI on benchmark datasets for graph classification and on the Elliptic Bitcoin dataset for blockchain fraud detection. The results demonstrate that ACLGMI consistently outperforms state-of-the-art unsupervised methods, achieving superior accuracy in graph classification and higher ROC-AUC scores in fraud detection. Our findings suggest that ACLGMI provides a generalizable and robust solution for structured data learning, with strong applications in blockchain security and beyond.</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

Beyond contrastive learning: adaptive graph representations with mutual information maximization for blockchain and structured data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月6日

查看全文
  • Yifeng Zhang,
  • Qianqian Ren,
  • Yourong Chen,
  • Meng Han

摘要

Graph representation learning has emerged as a powerful approach for modeling structured data across diverse domains, including social networks, biochemical interactions, and financial transaction systems. Existing contrastive learning frameworks have demonstrated strong performance in self-supervised graph learning; however, they often suffer from suboptimal augmentations, inadequate structural preservation, and a lack of adaptability to heterogeneous graph structures such as blockchain transaction networks. In this work, we propose ACLGMI (Adaptive Contrastive Learning with Graph Mutual Information Maximization), a novel framework that extends traditional contrastive learning by introducing an adaptive augmentation strategy and a mutual information maximization objective to enhance representation quality. ACLGMI dynamically adjusts augmentation strategies based on graph topology, ensuring that critical structural properties are preserved while improving robustness to noise and adversarial perturbations. Furthermore, we introduce a multi-level mutual information constraint that maximizes global and local consistency in learned representations, leading to improved performance across multiple downstream tasks. We evaluate ACLGMI on benchmark datasets for graph classification and on the Elliptic Bitcoin dataset for blockchain fraud detection. The results demonstrate that ACLGMI consistently outperforms state-of-the-art unsupervised methods, achieving superior accuracy in graph classification and higher ROC-AUC scores in fraud detection. Our findings suggest that ACLGMI provides a generalizable and robust solution for structured data learning, with strong applications in blockchain security and beyond.