A two-layer portfolio model based on machine learning for productivity prediction of garment industry

• Main Author: Zhou, Yiran
• Other Authors: Chen Songlin
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Manufacturing::Production management
• Online Access: https://hdl.handle.net/10356/166674
• Institution: Nanyang Technological University

Nowadays, garment industry is one of the most important sectors in this globalized industrial world. Being a labor-intensive industry, production heavily depends on workers' productivity in different factory departments. In this way, the common problem in this industry emerges: the actual productivity of employees fluctuates so that sometimes the target production goals can not be met. This gap between target productivity and real productivity usually causes obstacles to the flow of value chain, furthermore leads to economy and reputation loss to the companies. So, this issue is eagerly anticipated to be solved. This project aims to narrow the gap by predicting the productivity of employees in garment industry precisely. With more accurate prediction of the productivity, more reasonable production planning and scheduling can be made, and thus the gap between prediction and real productivity can be narrowed. To achieve this goal, a two-layer portfolio model based on machine learning regression algorithms and the stacking method of machine learning is proposed, which is able to predict the productivity of garment industry. Seven independent regression models, including ridge regression model, Bayesian linear regression model, polynomial regression model, SVM regression model, random forest regression model, GBDT regression model and neural network regression model, are used as base learners in the bottom layer while a random forest regression algorithm based generalizer is developed in the top layer, aiming to combine the independent regression models together to form one portfolio model. The portfolio model is trained and tested by a dataset collected from a garment factory in Bangladesh. Combining the seven regression models together, the portfolio model performs better than any of the independent models. With mean absolute error 0.0825, significantly lower than the MAE error baseline performance of 0.15, the experimental results proved the reliability of the proposed portfolio model. Such a model can be applied in industrial environments and help reduce the risks and uncertainties brought by the production links, and furthermore, minimize the loss of cost, maximize production efficiency and improve profits.