An interpretable predictive framework based on rules and machine learning: rule extraction, validation and adaptive integration with machine learning

Traditional knowledge extraction methods often rely on human expertise, which can be time-consuming and prone to cognitive biases. This work presents a comprehensive predictive framework that integrates rule extraction with machine learning (ML) to enhance knowledge discovery in materials science. We used subgroup discovery algorithms to extract rules based on their significance and categorized the datasets accordingly. These rules are considered implicit knowledge within the datasets. Then, the validation strategies are tailored to assess their effectiveness. Through case studies on high-entropy alloys and piezoelectric ceramics, we demonstrate that our rule-based subgrouping and rule embedding feature can significantly enhance model performance and interpretability compared to baseline models. It shows that the approach facilitates the interpretation of complex data, proving the potential of integrating interpretable rule-based insights with ML, and paving the way for future advancements in material design and predictive modeling.