Interactive Slice Visualization for Exploring Machine Learning Models

Machine learning models fit complex algorithms to arbitrarily large datasets. These are well known be high on performance and low interpretability. We use interactive visualization of slices predictor space address the interpretability deficit; in effect opening up black-box machine algorithms, for purpose interrogating, explaining, validating comparing model fits. Slices specified directly through interaction, or using various touring designed visit high-occupancy sections, regions where fits have interesting properties. The methods presented here implemented R package condvis2. Supplementary files this article available online.