Agnostic Explanation of Model Change based on Feature Importance

Abstract Explainable Artificial Intelligence (XAI) has mainly focused on static learning tasks so far. In this paper, we consider XAI in the context of online dynamic environments, such as from real-time data streams, where models are learned incrementally and continuously adapted over course time. More specifically, motivate problem explaining model change , i.e. difference between before after adaptation, instead themselves. regard, provide first efficient model-agnostic approach to dynamically detecting, quantifying, significant changes. Our is based an adaptation well-known Permutation Feature Importance (PFI) measure. It includes two hyperparameters that control sensitivity directly influence explanation frequency, a human user can adjust method individual requirements application needs. We assess validate our method’s efficacy illustrative synthetic streams with three popular classes.