Towards Evaluation of Explainable Artificial Intelligence in Streaming Data

This study introduces a method to assess the quality of Explainable Artificial Intelligence (XAI) algorithms in dynamic data streams, concentrating on the fidelity and stability of feature-importance and rule-based explanations. We employ XAI metrics, such as fidelity and Lipschitz Stability, to compare explainers between each other and introduce the Comparative Expert Stability Index (CESI) for benchmarking explainers against domain knowledge. We adopted the aforementioned metrics to the streaming data scenario and tested them in an unsupervised classification scenario with simulated distribution shifts as different classes. The necessity for adaptable explainers in complex scenarios, like failure detection is underscored, stressing the importance of continued research into versatile explanation techniques to enhance XAI system robustness and interpretability.