Unsupervised Risk Matrix Learning from Textual Data

• Grant number: 24/03755-7
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: May 01, 2024
• End date: April 30, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Ricardo Marcondes Marcacini
• Grantee: Lucas de Oliveira Ferreira
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil

Abstract

In this Scientific Initiation project, an approach is proposed for the automatic construction of risk matrices through machine learning methods. A risk matrix is a popular tool for project risk management, identifying potentially harmful events, their probability of occurrence, and their impact. Traditionally, the risk matrix is built with manual effort from domain experts, making the process costly and dependent on specific domain knowledge. Some machine learning initiatives for risk matrix construction have been proposed but still rely on a significant amount of labeled data, making the process also dependent on domain experts for data annotation. In this project, an alternative is proposed that combines (1) recent advances in Large Language Models (LLMs) to extract a set of events and perform an initial analysis of associated risks from representative domain texts, and (2) risk event modeling based on graphs to explore relationships between similar events and obtain final severity predictions with the support of Graph Neural Networks. Thus, as an expected result, the proposed approach enables the construction of the risk matrix without the need for labeled data. The obtained risk matrix can be presented to an organization or experts, who can make adjustments more cost-effectively.