• counterfactual
• feature importance
• model agnostic
• rule extraction

This thesis addresses the "black box" nature of artificial intelligence, particularly in critical applications, by enhancing model interpretability and transparency. Traditional feature importance methods like SHAP can be computationally intensive and sensitive to feature correlation, while counterfactual explanations typically focus on single-instance outcomes. The thesis seeks to improve upon these methods by optimizing the BoCSoR algorithm, which uses counterfactual explanations to extract global feature importance.

The study tests two new BoCSoR modifications—binary search in the feature space and feature-space pruning—against the original BoCSoR and methods like SHAP, DiCE, and Permutation Importance. Results show the binary search variant as the fastest, offering strong performance in capturing feature importance across the model. Additionally, the thesis proposes a new rule extraction approach based on counterfactual reasoning, benchmarking it against methods like RuleFit, to generate consultable rules that reveal the model’s decision-making logic.

This groundwork is essential for applying BoCSoR BinarySearch to the EXPERIENCE dataset, which focuses on creating personalized VR environments in healthcare, gaming, and education to enhance user engagement. By identifying significant features linked to mood disorders, the approach aims to yield interpretable, global feature importance and actionable rules that could inform treatment and understanding of depression, distinguishing features critical for both positive and negative diagnoses.