A Modified Ant Colony Optimization Bagging Framework for Software Defect Prediction

Software defects represent a persistent challenge in modern software development, frequently leading to system failures, elevated maintenance costs, and reduced software reliability. Conventional single-classifier defect prediction models often fail to capture the complex, non-linear relationships among software metrics, resulting in limited predictive accuracy and elevated false positive rates. This paper proposes MACO-BAG, a hybrid framework integrating Modified Ant Colony Optimization (MACO) with Bagging ensemble learning for software defect prediction. MACO incorporates an Elitism Mutation mechanism to overcome the premature convergence of standard ACO, enabling more precise and stable feature selection. The selected features train an ensemble of Decision Trees, SVMs, and Random Forests within a Bagging framework. Experiments on the NASA-MDP dataset (1,048 software modules) under four train–test splits demonstrate MACO-BAG achieving 98.12% accuracy, 97.89% precision, 99.00% sensitivity, 96.80% specificity, and a false positive rate of 3.20% in 17.58 seconds; consistently outperforming baseline Bagging (94.62%) and ACO-Bagging (96.42%).