Enhancing software effort estimation with random forest tuning and adaptive decision strategies

Area of Science:

• Computer Science
• Software Engineering
• Machine Learning

Background:

• Software Effort Estimation (SEE) is crucial for project management, resource allocation, and planning.
• Achieving precise SEE predictions remains a challenge in the software industry.
• Machine learning algorithms, particularly Random Forest (RF), show promise for improving SEE accuracy.

Purpose of the Study:

• To enhance the accuracy and interpretability of software effort estimation models.
• To introduce an Improved Adaptive Random Forest (IARF) model by integrating advanced techniques.
• To provide more trustworthy decision-making in software development projects.

Main Methods:

• Implemented an Improved Random Forest (IRF) with residual analysis, partial dependence plots, and feature engineering.
• Extended IRF to an adaptive model (IARF) using Bayesian Optimization with Deep Kernel Learning (BO-DKL) for hyperparameter tuning.
• Integrated Time-Series Residual Analysis for autocorrelation detection and Explainable AI (SHAP, LIME) for feature interpretability.

Main Results:

• The IARF model demonstrated significant improvements across various metrics compared to the standard Random Forest (RF) model.
• Average improvement ratios included: 18.5% on Mean Absolute Error (MAE), 20.3% on Root Mean Square Error (RMSE), 3.8% on R-Squared, and 5.4% on Mean Absolute Percentage Error (MAPE).
• A 7% reduction in Mean Absolute Scaled Error (MASE) and a 3-5% improvement in Prediction Interval Coverage Probability (PICP) were also observed.

Conclusions:

• The developed Improved Adaptive Random Forest (IARF) model substantially enhances software effort estimation accuracy.
• The combination of adaptive learning and explainability techniques leads to more reliable predictions.
• The findings support more trustworthy decision-making in software development project management.