Scholarlink

<p>Privacy constraints on industrial chain data pose significant challenges for effective risk assessment. In this paper, we propose a Graph Topology-aware Attention model for industrial chain Risk-Assessment task (GTARA). Specifically, we first construct semantic-specific meta-paths and employ a gated embedding mechanism to integrate and complete missing structural features of nodes, thereby enhancing the expressiveness of node representations. Subsequently, a risk injection mechanism is integrated into heterogeneous graph embedding learning, introducing adaptive perturbations to improve the robustness and discriminative power of node risk representations. Finally, we design a multi-task joint risk assessment method that incorporates attention pooling, enabling simultaneous optimization of node-level and graph-level risk evaluations to achieve a comprehensive multi-level assessment. We validate the proposed GTARA model on two real-world industrial datasets from the integrated circuit and electronic information domains. Results on five evaluation metrics, including AUC, F1-score, Accuracy, and additional performance indicators, demonstrate the effectiveness and superiority of the proposed method over existing baseline models.</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

Graph topology-aware attention for industrial chain risk assessment

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Pengzhao Sun,
  • Dongfen Li,
  • Chao Tang,
  • Fengli Zhang,
  • Ruijin Wang

摘要

Privacy constraints on industrial chain data pose significant challenges for effective risk assessment. In this paper, we propose a Graph Topology-aware Attention model for industrial chain Risk-Assessment task (GTARA). Specifically, we first construct semantic-specific meta-paths and employ a gated embedding mechanism to integrate and complete missing structural features of nodes, thereby enhancing the expressiveness of node representations. Subsequently, a risk injection mechanism is integrated into heterogeneous graph embedding learning, introducing adaptive perturbations to improve the robustness and discriminative power of node risk representations. Finally, we design a multi-task joint risk assessment method that incorporates attention pooling, enabling simultaneous optimization of node-level and graph-level risk evaluations to achieve a comprehensive multi-level assessment. We validate the proposed GTARA model on two real-world industrial datasets from the integrated circuit and electronic information domains. Results on five evaluation metrics, including AUC, F1-score, Accuracy, and additional performance indicators, demonstrate the effectiveness and superiority of the proposed method over existing baseline models.