Salon D: Generating comprehensive individual score explanations for segmented scorecard systems (FICO)
Salon D
Segmented scorecard systems are widely used in the credit risk scoring industry due to their ability to capture nuanced predictive patterns and behavioral heterogeneity (nonlinearities and interactions) in a transparent model architecture, based on intrinsically interpretable scorecards and segmentation rules. At the same time, these sophisticated glass-box models are on par with black-box machine learning models in terms of predictive performance in real-world applications. In this presentation, we will explore our latest research into developing individual (or ‘local’) score explanations for segmented scorecard systems that are based on segmentation trees defined by splitter features and rules. We will discuss an additive score decomposition approach that assigns score contributions to both scorecard features and splitter features, ensuring that individual score explanations are thorough, complete, and easy to comprehend. We will talk about the benefits of our novel approach, which employs Shapley values popularized in the interpretable machine learning area. In practical terms, these advancements may be applied to strategic segmentation, portfolio management, product and offer development, targeted messaging, and enhancing client engagement.
