Cost-sensitive learning and threshold-moving approach to improve industrial lots release process on imbalanced datasets

Abstract

With Industry 4.0, companies must manage massive and generally imbalanced datasets. In an automotive company, the lots release decision process must cope with this problem by combining data from different sources to determine if a selected group of products can be released to the customers. This work focuses on this process and aims to classify the occurrence of customer complaints with a conception, tune and evaluation of five ML algorithms, namely XGBoost (XGB), LightGBM (LGBM), CatBoost (CatB), Random Forest(RF) and a Decision Tree (DT), based on an imbalanced dataset of automatic production tests. We used a non-sampling approach to deal with the problem of imbalanced datasets by analyzing two different methods, cost-sensitive learning and threshold-moving. Regarding the obtained results, both methods showed an effective impact on boosting algorithms, whereas RF only showed improvements with threshold-moving. Also, considering both approaches, the best overall results were achieved by the threshold-moving method, where RF obtained the best outcome with a F1-Score value of 76.2%.