An Explainable Machine Learning and Optimization Framework for Profitability-Driven E-Waste Refurbishment

Abstract

Electronic waste (e-waste) refurbishment can extend device life, reduce material waste, and recover economic value, but operational decisions must balance uncertain resale outcomes against repair cost, labor, device condition, and limited resources.This paper presents an explainable machine-learning and optimization framework for profitability-driven refurbishment decision support. The study uses 51 operational refurbishment records containing device attributes, condition indicators, repair-related costs, labor effort, and resale outcomes. Logistic Regression, Random Forest, Support Vector Machine (SVM), and XGBoost are compared using stratified five-fold cross-validation. Feature-importance analysis is used to interpret the economic and condition-related drivers of predicted profitability, and a binary optimization model selects devices subject to budget and labor constraints. In the reported comparison, XGBoost achieved the highest F1-score (0.90), while Random Forest achieved the highest ROC-AUC (0.97). Under the evaluated resource-constrained scenario, optimization produced a total profit of $1,033 compared with $235 for a high-resale-value heuristic. Because the dataset contains only 51 devices from a limited operational setting, the findings should be interpreted as preliminary and require validation on larger and more diverse datasets.