Feature selection is a critical bottleneck in software defect prediction (SDP), where high-dimensional software metric datasets frequently contain noisy, redundant, or irrelevant attributes that degrade classifier performance. Although Ant Colony Optimization (ACO) is widely used for combinatorial feature selection, it suffers from premature convergence and search stagnation caused by excessive pheromone reinforcement of early high-quality solutions. This paper proposes Modified Ant Colony Optimization (MACO), an enhanced feature selection algorithm that integrates two complementary mechanisms: Elitism, which grants additional pheromone reinforcement to the globally best solutions, and Mutation, which introduces controlled stochastic diversity through probabilistic bit-flip operations on elite solutions. MACO is evaluated on the NASA Metrics Data Program (NASA-MDP) dataset comprising 1,048 software modules described by 23 metrics, using Decision Tree classifiers within a 5-fold cross-validation framework across four train–test splits (60–40, 70–30, 75–25, 80–20). Results demonstrate that MACO-based feature selection achieves 98.12% accuracy, 97.89% precision, 99.00% sensitivity, and 96.80% specificity with a false positive rate of just 3.20% outperforming standard ACO (96.42% accuracy, 5.33% FPR) across all configurations. Convergence analysis further confirms that MACO reaches near-optimal fitness within 15–20 iterations, compared to approximately 40 for standard ACO, while exhibiting markedly reduced oscillation. These findings establish MACO as a powerful, computationally efficient feature selection mechanism for high-dimensional software defect datasets.

Keywords: Feature Selection; Ant Colony Optimization; Elitism Mutation; Software Defect Prediction; Pheromone Update; Premature Convergence; NASA-MDP; Metaheuristic Optimization.