Abstract
Software engineering companies strive to improve software quality by predicting software defects-prone modules. Although various data mining methods have been developed, unstable accuracy rates are still critical issues owing to the imbalanced nature and high dimensionality of software defect datasets. To deal with this issue, we propose a spotted hyena, a novel meta-heuristic optimization algorithm for predicting software defects. Support and confidence in classification rules are the basis of a multi-objective fitness function that assists the spotted hyena algorithm in serving as a classifier by finding the fittest classification or standard rules among individuals. Experiments were conducted on four NASA software datasets, JM1, KC2, KC1, and PC3. The spotted hyena classifier provides an accuracy of 85.2, 84, 89.6, and 81.8%, respectively, for these datasets. These accuracy rates are better than those achieved using other popular data mining techniques. We also discuss other classification measures in connection with the experimental results, such as precision, recall, and confusion matrices, in connection with the experimental results. Moreover, the Gaussian mixture model is used to study the uncertainty quantification of the proposed classifier. The study proved the feasible performance of the spotted hyena classifier in four different case studies.