Software defect prediction using support vector machines with adaptive particle swarm optimization algorithm

Software defect prediction (SDP) is significant, because it contributes to allocating test resource reasonably, increasing test efficiency and improving software quality by classifying a specified software module as defect-prone or no defect-prone. Usually, feature selection method will be used to select the essential software metrics as the input variables of SDP model. However, the selected software metrics often vary obviously with the number of training sample even as the same dataset. We think this is unreasonable. This paper proposed an easy and practicable approach to determine the metrics needed by modeling and a hybrid SDP model. By regarding the software defect prediction problem as a binary classification problem, a classification model is established by support vector machines (SVM) and an improved particle swarm optimization (PSO) algorithm by adopting adaptive inertial weight and adaptive mutation (APSO). Case study was performed based on four public datasets from PROMISE software engineering repository and the performance was evaluated by comparing with SVM and PSO-SVM in terms of ROC curve and AUC value. Experimental results showed that the proposed model perform better comparing with the other models in general.