Visual Explanations and Perturbation-Based Fidelity Metrics for Feature-Based Models

This work introduces an enhanced methodology in the domain of eXplainable Artificial Intelligence (XAI) for visualizing local explanations of black-box, feature-based models, such as LIME and SHAP, enabling both domain experts and non-specialists to identify the segments of Time Series (TS) data that are significant for machine learning model interpretations across classes. By applying this methodology to electrocardiogram (ECG) data for anomaly detection, distinguishing between healthy and abnormal segments, we demonstrate its applicability not only in healthcare diagnostics but also in predictive maintenance scenarios. Central to our contribution is the development of the AUC Perturbational Accuracy Loss metric (AUC-PALM), which facilitates the comparison of explainer fidelity across different models. We advance the field by evaluating various perturbation methods, demonstrating that perturbations centered on time series prototypes and those proportional to feature importance outperform others by offering a more distinct comparison of explainer fidelity with the underlying black-box model. This work lays the groundwork for broader application and understanding of XAI in critical decision-making processes.