Ethics-Driven Transparency in Customer Analytics: A Fuzzy Cognitive Mapping Approach to Explainable Churn Prediction in E-Commerce

Abstract

The opacity of high-performing machine learning (ML) models in customer analytics creates ethical and regulatory challenges, as organizations struggle to balance predictive accuracy with the explainability mandated by international data protection and artificial intelligence regulations. This research develops and validates a hybrid fuzzy cognitive map–machine learning (FCM-ML) framework addressing the performance-interpretability trade-off in customer churn prediction. The study focuses on Moroccan e-commerce, a rapidly digitalizing context with significant customer attrition challenges, which makes transparent analytics solutions particularly relevant for both ethical and operational reasons. The empirical investigation analyzes 50,000 customers from the MarocShop platform over a 24-month period (January 2023 – December 2024), capturing comprehensive behavioral, transactional, engagement, and service metrics. The methodology introduces a novel automated algorithm extracting FCMs through principal component analysis (PCA), Granger causality testing, and normalized mutual information (NMI) quantification, integrated with ensemble ML (random forest, XGBoost, LightGBM). Results confirm the research hypothesis that transparency and performance are simultaneously achievable: the hybrid framework attains 89.3–90.1% accuracy, matching state-of-the-art black-box approaches, while maintaining 91% causal transparency via the Causal Transparency Index (CTI) – a 118% explainability improvement over baseline models. Three causal pathways were identified: behavioral disengagement (38%), dissatisfaction escalation (26%), and transactional friction (22%), with a quantified business impact of +4.9% campaign return on investment (ROI) (+79,800 MAD per 2,000 customers), scaling to +17.9 million MAD annually. These results advance ethical business practices by demonstrating that algorithmic transparency and predictive performance are compatible objectives.