Impact of Resample and Ensembles on Software Maintainability Prediction

Software maintainability prediction (SMP) ensures long-term software sustainability and minimizes development costs.However, data imbalance often leads to biased classifications and inaccurate predictions.This research proposes a comprehensive approach to address SMP data imbalance and accuracy issues.The method utilizes Resample, Class Balancer, and Spread Subsample techniques to balance a large, unbalanced dataset comprising seven software packages: LUCENE, JDT, PDE, EQUINOX, MYLENE, UIMS, and QUES.We evaluated six machine learning models for the training of SMP models: Support Vector Machine, Decision Tree, Nave Bayes, k-nearest Neighbour, and two ensemble methods (Bagged Tree and Boosted Tree).TOPSIS analysis identifies the Bagged Tree with Resample as the best-performing model, achieving a maximum accuracy of 95.7%.Shapley Additive Explanations and Local Interpretable Model-Agnostic Explanations techniques analysed the interpretability of the best SMP model, revealing key predictors influencing software maintainability.Finally, the Mean Magnitude of Relative Error (MMRE) demonstrates that the proposed model outperforms existing studies.This work confirms its effectiveness, reliability, and practical use in software engineering.