Failing Conceptually: Concept-Based Explanations of Dataset Shift
This addresses the need for actionable insights in shift detection for practitioners, though it is incremental as it builds on existing detection techniques by adding explainability.
The paper tackles the problem of machine learning models failing under data distribution shifts by introducing Concept Bottleneck Shift Detection (CBSD), a method that explains shifts by identifying affected human-understandable concepts, achieving higher detection accuracy than state-of-the-art methods in case studies on dSprites and 3dshapes.
Despite their remarkable performance on a wide range of visual tasks, machine learning technologies often succumb to data distribution shifts. Consequently, a range of recent work explores techniques for detecting these shifts. Unfortunately, current techniques offer no explanations about what triggers the detection of shifts, thus limiting their utility to provide actionable insights. In this work, we present Concept Bottleneck Shift Detection (CBSD): a novel explainable shift detection method. CBSD provides explanations by identifying and ranking the degree to which high-level human-understandable concepts are affected by shifts. Using two case studies (dSprites and 3dshapes), we demonstrate how CBSD can accurately detect underlying concepts that are affected by shifts and achieve higher detection accuracy compared to state-of-the-art shift detection methods.