Klasifikasi Kecacatan Produksi Manufaktur Menggunakan Ensemble Learning: Perbandingan XGBoost, LightGBM, dan Random Forest dengan SHAP Explainability

Abstract

Production defects represent a critical challenge in modern manufacturing industries, directly impacting operational efficiency, production costs, and customer satisfaction. This study proposes an Ensemble Learning-based approach to classify production defect status using the Predicting Manufacturing Defects dataset from Kaggle (3,240 records, 16 features). Three state-of-the-art Ensemble Learning algorithms, XGBoost, LightGBM, and Random Forest, were comprehensively evaluated against Support Vector Machine (SVM) and K-Nearest Neighbor (KNN) as baseline models. Significant class imbalance (84.04% High Defects vs. 15.96% Low Defects) was addressed using Synthetic Minority Over-sampling Technique (SMOTE). Evaluation employed Accuracy, Precision, Recall, F1-Score, AUC-ROC, and 5-fold Cross-Validation. SHapley Additive exPlanations (SHAP) was applied to enhance model interpretability and identify the most influential features. Results show Random Forest achieved the highest accuracy of 94.75% with F1-Score 94.49%, while LightGBM performed best in 5-fold Cross-Validation with mean F1-Score of 95.92% ± 1.07%. SHAP analysis revealed that MaintenanceHours, DefectRate, and QualityScore are the three most dominant factors in determining production defect status.