Unsupervised software defect prediction using signed Laplacian-based spectral classifier

The lack of training dataset availability is the most popular issue in the software defect prediction, especially when dealing with new project development. Adopting training dataset from other software projects probably will not be the best solution because of the software metrics heterogeneity issues across projects. Unsupervised approaches have been proposed to address this issue, where the software prediction model is built without training dataset. Spectral classifier is one of these unsupervised approaches that has been applied successfully to address the lack of training dataset. However, this method leaves an issue when the dataset does not meet the requirement of nonnegative Laplacian assumption. This case would be occurred if there were nonnegative values of the adjacency matrix. It is well known that spectral classifier works with the Laplacian matrix, where the Laplacian matrix is constructed by adjacency matrix. In this paper, the signed Laplacian-based spectral classifier is proposed to solve the negative values problem in the adjacency matrix by converting the negative values into absolute values. The experimental results show that the proposed method could improve the performance of unsupervised classifiers compared to the unsigned Laplacian-based spectral classifier method. Hence, the proposed method is strongly suggested as unsupervised software defects prediction for the software projects that have no historical software dataset.