DDImage: an image reduction based approach for automatically explaining black-box classifiers

Due to the prevalent application of machine learning (ML) techniques and the intrinsic black-box nature of ML models, the need for good explanations that are sufficient and necessary towards locally interpreting a model’s prediction has been well recognized and emphasized. Existing explanation approaches, however, favor either the sufficiency or necessity. To fill this gap, in this paper, we propose an approach for generating local explanations that are both sufficient and necessary. Our approach, DDImage, automatically produces local explanations for ML-based image classifiers in a post-hoc way. The core idea behind DDImage is to discover an appropriate explanation by debugging the given input image via a series of image reductions, with respect to the sufficiency and necessity properties. Evaluation of DDImage using publicly available datasets and popular classification models reveals its effectiveness and efficiency. Compared with three state-of-the-art approaches, DDImage demonstrates a superior performance in producing small-sized explanations preserving both sufficiency and necessity, and it also shows promising stability and efficiency. We also identify the impact of segmentation granularity, reveal the performance variance for different target models, and further show that our approach is applicable across different problem domains.