Hybrid machine learning approach for scrap prediction: An xgboost and catboost integration

Accurate prediction of scrap in industrial production is critical for reducing waste, lowering costs, and sustaining product quality, yet it is challenging due to high dimensional, mixed-type datasets and complex nonlinear relationships between process variables and quality outcomes. This thesis develops an end-to-end data preparation and modelling pipeline for a large real world manufacturing dataset that includes domain informed feature selection, handling of missing values, and Local Outlier Factor (LOF) based outlier suppression for improved interpretability. At its core, the study introduces an original hybrid prediction framework, referred to as the proposed method, in which tuned CatBoost and XGBoost models are combined through a metaheuristic-optimised, confidence-aware blending mechanism equipped with a backup model for low confidence instances. Genetic Algorithms are used to tune key hyperparameters of the base learners, while Particle Swarm Optimization configures the blending stage so that the contribution of each base model is adapted on a per instance basis. The proposed method is evaluated against a broad set of tuned competitors including classical regressors, tree based ensembles, stand-alone gradient boosting models, a neural network, and a fixed weighted average baseline using multiple regression metrics (MAE, MSE, RMSE, R^2, RAE, RSE). Results show that the proposed method achieves the lowest error values and the highest explanatory power (R^2≈0.84), and that its predictions closely follow the temporal evolution of scrap amounts in time series analyses. Overall, the thesis demonstrates that feature aware, metaheuristic calibrated blending of CatBoost and XGBoost, embedded in a robust data preparation pipeline, offers a scalable and interpretable solution for scrap prediction in mixed-type industrial datasets.