Software Defect Prediction Using Dissimilarity Measures

In order to improve the accuracy of software defect prediction, a novel method based on dissimilarity measures is proposed. Different from traditional predicting methods based on feature space, we solve the problem in dissimilarity space. First the new unit features in dissimilarity space are obtained by measuring the dissimilarity between the initial units and prototypes. Then proper classifier is chosen to complete prediction. By prototype selecting, we can reduce the dimension of units´ features and the computational complexity of prediction. The empirical results in the NASA database KC2 and CM1 show that the prediction accuracies of KNN, Bayes, and SVM classifier in dissimilarity space are higher than that of feature space from 1.86% to 9.39%. Also the computational complexities reduce from 18% to 67%.