BARBiE: An Associative Rule-Based Interactive Framework for Explaining Black-Box Model

Post-hoc explainable artificial intelligence is often provided as a product, typically in the form of static explanation such as a feature-importance ranking or a local surrogate explanation. In contrast, real-world decision workflows demand explanation as a process, characterized by interactivity in which users explore the decision output with what-if questions to develop understanding and trust. Existing explainers are often static, and their output is sensitive to how local samples around the instance are selected. Although rule-based local surrogates can expose feature interactions, user edits often require repeated resampling and retraining, limiting their usability for real-time what-if analysis. To address these gaps, we introduce BARBiE, a model-agnostic framework for instance-level explanation that integrates an association-rule surrogate with an interactive interface. For a given query instance, BARBiE constructs an instance-centered neighborhood, queries the black-box model for labels, and trains a compact association-rule surrogate. Explanations are provided only when the surrogate output matches the black-box decision for the query instance. BARBiE presents IF–THEN rules with support, confidence, and a p-value from Fisher’s exact test. In addition, BARBiE computes rule-grounded, signed feature importance by aggregating instance-aware contributions from the rule base. Importantly, BARBiE supports quick what-if analysis without resamples and retraining the surrogate model. Across four tabular datasets and a user study, we evaluated BARBiE against LIME, SHAP, and BARBE using user ratings of informativeness, understandability, trustworthiness, and satisfaction. Across tasks, BARBiE consistently received higher ratings than the baselines, providing supports that process-centric interactive explanations improve informativeness and understandability and contribute to higher trust and user satisfaction.