Explain, Embed, Retrieve, and Reason (E2R2): A SHAP-Informed LLM Framework for Decision Support

This paper introduces E2R2 (Explain, Embed, Retrieve, and Reason), a framework that combines SHAP-based feature attribution, case-based retrieval, and GPT-driven reasoning for explainable classification. Applied to student attrition prediction, E2R2 achieved 89.2% accuracy and 84.1 F1 on a 500-case holdout set, comparable to Decision Tree and Random Forest baselines while offering higher recall and balanced precision–recall. Validation showed 94% consistency across GPT sessions and resilience to incomplete data (accuracy = 86.6% under 10% feature dropout). Beyond predictive accuracy, E2R2 generates SHAP-grounded, peer-informed narratives that improve cognitive accessibility. Although demonstrated in higher education, the architecture is domain-agnostic and adaptable to fields such as healthcare or finance. By extending feature attributions into context-aware explanations, E2R2 exemplifies the design of next-generation decision support systems that combine analytic precision with interpretability.