A Multi-Tier Framework for Ranking Determinants in Quantitative Modeling using xAI and LLM-Orchestrated Argumentation

In the era of e-commerce, understanding consumer behavior is critical, but traditional linear models often fail to capture the complex, nonlinear relationships that drive real-world decisions. While nonlinear AI models offer superior predictive power, their “black-box” nature creates a critical gap in interpretability, hindering managerial trust and actionable insights. Existing explainable AI (XAI) techniques often reveal why a model makes a prediction but fall short in providing practical, minimal-effort recommendations on how to achieve a desired outcome. To bridge this gap, we propose a novel, multi-tier XAI framework. The core of our framework is SHAP-guided DiCE, a new method that intelligently integrates global model insights from SHAP into the counterfactual generation process. Applying this to a consumer privacy dataset ($\mathrm{N}=467$), we demonstrate that our variant significantly outperforms the baseline, generating valid counterfactuals requiring minimal changes-often only a single feature modification compared to 4-5-while producing a more diverse set of actionable recommendations (diversity score of 2.24 vs. 0). The framework culminates in a humancentered, LLM-orchestrated argumentation layer, which synthesizes these quantitative outputs into structured, contestable narratives, transforming static interpretations into a dynamic, data-driven dialogue tailored for decision-makers. This work moves beyond simple explanation towards explainable action, providing a transparent framework to identify the most efficient levers for influencing behavioral outcomes.