Abordagem de CRM orientada por aprendizado de máquina para identificar evasão de cooperados em uma cooperativa agroindustrial brasileira: estudo de caso prático

Abstract

This study addresses member churn in a Brazilian agro-industrial cooperative by operationalizing a leakage-aware, governance-aligned machine-learning protocol within the organization’s Customer Relationship Management (CRM) system. Using real-world CRM data under confidentiality constraints, we followed a KDD-based workflow. This workflow includes: (i) multisource integration; (ii) targeted preprocessing with explicit handling of severe class imbalance via undersampling; (iii) a unified validation scheme with stratified cross-validation, hyperparameter search, and controlled AutoML benchmarking; (iv) comparison of tabular learners (Random Forest, XGBoost, and Support Vector Machine) and a voting ensemble; and (v) SHAP-based explainability to support transparent decision-making. Class rebalancing substantially improved minority-class performance; for instance, the “Inactive” recall increased from 0.27 to 0.74 with SVM. Across ten folds, AutoML achieved competitive mean ROC-AUC (0.8844), followed by XGBoost (0.8690) and Random Forest (0.8660); global metrics supported operational feasibility (accuracy 0.79-0.80; ROC-AUC up to 0.8876), while the ensemble delivered comparable discrimination (ROC-AUC 0.8845) with a modest precision gain. SHAP analyses yielded business-coherent drivers and enabled actionable, instance-level communication in the CRM. The resulting microservices-based module exposes ranked churn propensities and explanations in dashboards for risk stratification and prioritization of retention actions. Overall, the work provides an interpretable, reproducible, and production-ready methodological blueprint for predictive CRM in seasonal cooperative environments under governance and confidentiality constraints.