Fast and More Robust Permutation Importance for Black-box Model Transparency on Imbalanced and High-Dimensional Data
By Michal Zak
In the first part of this blog, Basic Concepts and Techniques of AI Model Transparency, we reviewed a few common techniques for AI model transparency such as linear coefficients, local linear approximation, and permutation importance. In particular, the permutation importance is applicable to any black-box models, any accuracy/error functions, and more robust against high-dimensional data (because it handles each feature one by one rather than all features at the same time). One of the drawbacks of the permutation importance is its high computation cost. We have to repeat the evaluation process by (the number of features) * (the number of random shuffling to repeat) * (the number of models). To reduce the computation time, a common approach is to apply downsampling that works well when the positive and negative classes are balanced. However, such naive downsampling makes permutation importance extremely unreliable. Permutation Importance Under Class Imbalance Let us first see…