Comparing Stacking Ensemble and Deep Learning for Software Project Effort Estimation

This study focuses on improving the accuracy of effort estimation by employing ensemble, deep learning, and transfer learning techniques. An ensemble approach is utilized, incorporating XGBoost, Random Forest, and Histogram Gradient Boost as generators to enhance predictive capabilities. The performance of the ensemble method is compared against both the deep learning approach and the PFA-IFPUG technique. Statistical criteria including MAE, SA, MMRE, PRED(0.25), MBRE, MIBRE, and relevant information related to MMRE and PRED(0.25) are employed for evaluation. The results demonstrate that combining regression models with Random Forest as the final regressor and XGBoost and Histogram Gradient Boost as prior generators yields more accurate effort estimation than other combinations. Furthermore, the findings highlight the potential of transfer learning in the deep learning method, which exhibits superior performance over the ensemble approach. This approach leverages pre-trained models and continuously improves performance by training on new datasets, providing valuable insights for cross-company and cross-time effort estimation problems. The ISBSG dataset is used to build the pre-trained model, and the inductive transfer learning approach is verified based on the Desharnais, Albrecht, Kitchenham, and China datasets. The study underscores the significance of transfer learning and the integration of domain-specific knowledge from existing models to enhance the performance of new models, thereby improving accuracy, reducing errors, and enhancing predictive capabilities in effort estimation.