Unsupervised software defect prediction using median absolute deviation threshold based spectral classifier on signed Laplacian matrix

Abstract

Abstract Area of interest The trend of current software inevitably leads to the big data era. There are much of large software developed from hundreds to thousands of modules. In software development projects, finding the defect proneness manually on each module in large software dataset is probably inefficient in resources. In this task, the use of a software defect prediction model becomes a popular solution with much more cost-effective rather than manual reviews. This study presents a specific machine learning algorithm, which is the spectral classifier, to develop a software defect prediction model using unsupervised learning approach. Background and objective The spectral classifier has been successfully used in software defect prediction because of its reliability to consider the similarities between software entities. However, there are conditional issues when it uses the zero value as partitioning threshold. The classifier will produce the predominantly cluster when the eigenvector values are mostly positives. Besides, it will also generate low clusters compactness when the eigenvector contains outliers. The objective of this study is mainly to propose an alternative partitioning threshold in dealing with the zero threshold issues. Generally, the proposed method is expected to improve the spectral classifier based software defect prediction performances. Methods This study proposes the median absolute deviation threshold based spectral classifier to carry out the zero value threshold issues. The proposed method considers the eigenvector values dispersion measure as the new partitioning threshold, rather than using a central tendency measure (e.g., zero, mean, median). The baseline method of this study is the zero value threshold based spectral classifier. Both methods are performed on the signed Laplacian matrix to meet the non-negative Laplacian graph assumption. For classification, the heuristic row sum method is used to assign the entity class as the prediction label. Results and conclusion In terms of clustering, the proposed method can produce better cluster memberships that affect the cluster compactness and the classifier performances improvement. The cluster compactness average of both the proposed and baseline methods are 1.4 DBI and 1.8 DBI, respectively. In classification performance, the proposed method performs better accuracy with lower error rates than the baseline method. The proposed method also has high precision but low in the recall, which means that the proposed method can detect the software defect more precisely, although in the small number in detection. The proposed method has the accuracy, precision, recall, and error rates with average values of 0.79, 0.84, 0.72, and 0.21, respectively. While the baseline method has the accuracy, precision, recall, and error rates with average values of 0.74, 0.74, 0.89, and 0.26, respectively. Based on those results, the proposed method able to provide a viable solution to address the zero threshold issues in the spectral classifier. Hence, this study concludes that the use of the median absolute deviation threshold can improve the spectral based unsupervised software defect prediction method.