EOELM: An Explainable Optimized Ensemble Learning Model for Intelligent Software Cost Estimation

Abstract

Accurate early cost estimation of software is critical to effective project planning, budgeting, resource allocation and risk management. However, the traditional algorithmic models and most of the contemporary machine learning models are restricted in terms of prediction power, rely on manually set hyperparameters and do not provide model interpretability, which limits their practical application in software project management. In this research, to solve these problems for intelligent software cost estimate, an Explainable Optimized Ensemble learning Model (EOELM) is proposed. The proposed approach is a single estimate model which includes intelligent data preparation, hybrid feature selection, hyperparameter optimization based on Grey Wolf Optimizer (GWO), weighted stacking ensemble learning and Explainable Artificial Intelligence (XAI). Datasets of software projects are pre-processed using missing value imputation, outlier reduction, normalization and feature engineering. A hybrid feature selection method based on Correlation Analysis, Mutual Information, Recursive Feature Elimination and the Boruta algorithm is used to discover the most relevant software cost drivers. Then, the hyperparameters of four complimentary base learners, eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Random Forest (RF), and Artificial Neural Network (ANN), are fine-tuned by GWO, and their predictions are combined via a weighted stacking ensemble for improving the accuracy and robustness of the estimate. SHAP and LIME enhance the interpretability of the model and explain the software cost estimates at global and local level. The suggested system is validated on benchmark data sets as NASA93, COCOMO81, Desharnais, Maxwell and ISBSG using 10 fold cross validation. The performance evaluation is done using MAE, RMSE, MMRE, MdMRE, Pred(25) and R² and statistical significance of the findings is checked using Wilcoxon signed-rank and Friedman tests. The experimental findings show that EOELM is superior than the current software cost estimating approaches in prediction accuracy, generalization ability and interpretability. The framework presented is a dependable and transparent decision support system for the intelligent planning and cost estimation of software projects.