COSHAP: An Explainable AI Framework Integrating Conformal Prediction and SHAP for Uncertainty-Aware Consumer Segmentation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Data-driven decision-making in digital marketing is often constrained by the lack of transparency and uncertainty estimation in conventional machine learning models. This study proposes COSHAP (Conformal Segmentation with Hierarchical Attribution Paths), a unified framework that integrates Conformal Prediction (CP) and SHAP to model uncertainty while enhancing interpretability in customer segmentation. Using the Online Retail II dataset, segmentation was performed based on Recency, Frequency, Monetary (RFM), Tenure, and Average Basket Size features using K-Means, which was subsequently approximated using supervised models (Random Forest, XGBoost, Naïve Bayes, and SVM). CP was applied to generate prediction sets with guaranteed statistical coverage, while SHAP analysis was extended by conditioning feature attributions on prediction set membership through Hierarchical Attribution Paths. Experimental results demonstrated that the XGBoost model achieved the highest approximation accuracy of 97.36%. At a 95% confidence level, the framework achieved empirical coverage consistent with the statistical target for the ensemble models, with an average prediction set size close to 1.0, indicating high certainty in the majority of cases. However, a subset of customers at decision boundaries yielded multi-label predictions, accurately reflecting behavioral ambiguity. SHAP analysis identified Recency and Frequency as the primary drivers of segmentation, whereas COSHAP successfully uncovered the push-pull feature dynamics that trigger prediction uncertainty. The COSHAP framework effectively transforms point-based segmentation into a reliable, calibrated, and transparent decision-support system.