Abstract: Aiming at the complexity and imbalance of data dimensions in software defect prediction (SDP), a software defect prediction method based on the surrogate-assisted multi-objective firefly algorithm (SMO-MSFFA) is proposed. The proposed method employs the multi-group strategy firefly algorithm (MSFFA) to take minimizing the feature selection ratio and maximizing the the model evaluation AUC value as the two objective functions. Random forest (RF), support vector machine (SVM), and K-nearest neighbor classification algorithm (KNN) are used as the classifiers to construct the software defect prediction models. Considering the computational complexity of the evolutionary algorithm, the embedded surrogate-assisted model completes the calculation of partial individual evaluation function offline to reduce the computational cost. Experiments on PC1, KC1, and MC1 of NASA datasets verify that, compared with NSGA-II, our method increases the model evaluation AUC value by 0.17 on PC1, decreases it by 0.01 on KC1, and increases it by 0.09 on MC1, decreases the average feature selection ratio by 0.08, 0.17, and 0.05 on PC1, KC1, and MC1 respectively, and increases the computational time by 131 seconds, and decreases the time by 199 seconds and 431 seconds on KC1 and MC1 respectively. Experimental results show that the proposed method has obvious advantages in improving the model performance, reducing the feature selection ratio and reducing the computational time. 

Key words: software defect prediction, machine learning, multi-objective, firefly algorithm, surro-gate-assisted model