Scalable black-box model explainability through low-dimensional visualizations

Abstract

Two methods are proposed to provide visual intuitive explanations for how black-box models work. The first is a projection pursuit-based method that seeks to provide data-point specific explanations. The second is a generalized additive model approach that seeks to explain the model on a more holistic level, enabling users to visualize the contributions across all features at once. Both models incorporate visual and interactive elements designed to create an intuitive understanding of both the logic and limits of the model. Both explanation systems are designed to scale well to large datasets with many data points and many features.