Globally-Consistent Rule-Based Summary-Explanations for Machine Learning Models: Application to Credit-Risk Evaluation
Cynthia Rudin, Yaron Shaposhnik; 24(16):1−44, 2023.
Abstract
This study presents a novel method for comprehending specific predictions made by global predictive models through the creation of local models tailored to each individual observation, which are commonly referred to as “explanations” in the literature. Unlike existing approaches that approximate global models in the vicinity of specific observations to explain them, our method constructs globally-consistent models that align with the predictions made by the global model on historical data. We specifically focus on rule-based models, also known as association rules or conjunctions of predicates, which are widely used and interpretable. We develop several algorithms to extract these rules from discrete and continuous datasets and analyze their theoretical properties. Furthermore, we apply these algorithms to various credit-risk models trained on the Explainable Machine Learning Challenge data from FICO, demonstrating that our approach can produce concise summary-explanations of these models within seconds. Our method is model-agnostic, meaning it can be utilized to explain any predictive model, and it solves a minimum set cover problem to construct its summaries.
[abs]
[code]