Exploring stacking methods for software effort estimation with hyperparameter tuning

Maryam Hassanali

doi:10.1007/s10586-024-04876-8

Exploring stacking methods for software effort estimation with hyperparameter tuning

2025-02-25

0

PRE

AI

M

Maryam Hassanali

M

Mohammadreza Soltanaghaei *

T

Taghi Javdani Gandomani

F

Farsad Zamani Boroujeni

DOI：10.1007/s10586-024-04876-8

原文链接

原文求助

摘要

En 中文

This study explores the use of stacking ensemble methods for software effort estimation, leveraging the ISBSG dataset. The research compares individual machine learning models such as Support Vector Regression (SVR), Random Forest, AdaBoost, LASSO, and neural networks with stacked models over two phases. In Phase 1, five stacking models were developed, using SVR, Random Forest, AdaBoost, LASSO, and a neural network as base models. In Phase 2, four stacking models incorporated two neural networks as base models combined with different metamodels. The results indicate that stacking models with SVR as the metamodel outperformed individual models and other ensemble methods, especially in terms of MAE, MSE, MMRE, and MdMRE evaluation metrics. Friedman Test and Nemenyi Post-Hoc Test validated the results. These findings suggest that using SVR as the metamodel leads to improved accuracy and reliability in effort estimation, offering significant advantages for project management by enhancing predictive precision and optimizing resource allocation.

Keyword:

Software effort estimation

Stacked generalization

Stacked ensemble learning

SVR

Random forest

AdaBoost

Exploring stacking methods for software effort estimation with hyperparameter tuning

摘要

期刊

机构