Scholarlink

<p>Risk analysis is a determining factor for every project, aiming to assess existing risks with the goal of minimizing financial and human losses. A well-established method for evaluating risks is the risk matrix, which comprises two dimensions: the probability and severity of events. Current machine learning approaches for constructing risk matrices require hundreds or thousands of manually labeled examples by domain experts and typically address only one dimension of the matrix. This work introduces RMGNN (Risk Matrix Generation with Graph Neural Networks), a semi-supervised method that constructs complete risk matrices using only five percent labeled data. RMGNN builds a unified graph structure where events are connected based on textual similarity, then extracts both dimensions from this structure: occurrence probability is estimated through graph centrality measures, while event severity is predicted through semi-supervised label propagation using a graph neural network. For risk managers, this approach reduces expert annotation costs by ninety-five percent and enables building customized risk matrices in days rather than weeks. The method also supports cross-domain knowledge transfer, allowing organizations to leverage labeled incidents from one domain to assess risks in related domains with limited historical data. Experimental evaluation on over ten thousand consumer product incidents demonstrates that RMGNN achieves eighty-one percent accuracy and seventy-four percent F1-score, significantly outperforming traditional semi-supervised baselines while requiring substantially less labeled training data.</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

Learning risk matrix from textual data through graph neural networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月5日

查看全文
  • Matheus Yasuo Ribeiro Utino,
  • Vinícius Teixeira Carvalho,
  • Ricardo Marcondes Marcacini

摘要

Risk analysis is a determining factor for every project, aiming to assess existing risks with the goal of minimizing financial and human losses. A well-established method for evaluating risks is the risk matrix, which comprises two dimensions: the probability and severity of events. Current machine learning approaches for constructing risk matrices require hundreds or thousands of manually labeled examples by domain experts and typically address only one dimension of the matrix. This work introduces RMGNN (Risk Matrix Generation with Graph Neural Networks), a semi-supervised method that constructs complete risk matrices using only five percent labeled data. RMGNN builds a unified graph structure where events are connected based on textual similarity, then extracts both dimensions from this structure: occurrence probability is estimated through graph centrality measures, while event severity is predicted through semi-supervised label propagation using a graph neural network. For risk managers, this approach reduces expert annotation costs by ninety-five percent and enables building customized risk matrices in days rather than weeks. The method also supports cross-domain knowledge transfer, allowing organizations to leverage labeled incidents from one domain to assess risks in related domains with limited historical data. Experimental evaluation on over ten thousand consumer product incidents demonstrates that RMGNN achieves eighty-one percent accuracy and seventy-four percent F1-score, significantly outperforming traditional semi-supervised baselines while requiring substantially less labeled training data.