Software Development Cost Prediction Using Ensemble Learning Techniques

Cost estimation of software is an important task and should be done to complete the software project successfully. Traditional approaches of software cost estimation suffer from subjectivity and additional cost. To overcome these limitations with the traditional methods, the field of Machine learning has been incorporated as a promising method for the problem of software cost estimation. This paper explores using ensemble learning to improve cost predictions on two public benchmark datasets, Albrecht and Maxwell. Sequential Forward Feature Selection (SFS) is applied to pick the most relevant features and simplify models. Four ensemble methods ("Bagging", "AdaBoost", "XGBoost", and "CatBoost") have been implemented and evaluated on the benchmark Albrecht and Maxwell datasets. To avoid overfitting and to assess the model properly, the 10-fold cross-validation strategy was used. The results indicate that model performance is dependent on the dataset. CatBoost performed best on the Albrecht dataset, achieving the highest R² score of 0.943, while AdaBoost achieved the best performance on the Maxwell dataset, with an R² score of 0.735. These findings demonstrate the effectiveness of ensemble learning methods in improving the predictive accuracy of software cost estimation models.