Software defect detection based on selected complexity metrics using fuzzy association rule mining and defective module oversampling

Software defect is a major problem in software development. The cost of software development will be minimized when the software defects are detected earlier. Complexity metric is a mathematic calculation to calculate code complexity. It could be used to consider software defect detection. But, not all of complexity metrics influent on the occurrence of software defect, therefore it needs feature selection to select the most influent complexity metrics. Correlation-based Feature Selection (CFS) is used for selecting the most influent complexity metrics. This study conducted experiments on NASA Metric Data Program (MDP) datasets. NASA MDP contains software defect history logs based on several complexity metrics. But, there is an imbalanced distribution of defective and not defective modules in NASA MDP. The distribution of defective modules is less than not defective modules. It can reduce software defect detection performance. The distribution of defective module need to be reproduced. In this study, Synthetic Minority Oversampling Technique (SMOTE) is used to balance the distribution between defective and not defective modules. Software defect detection using Fuzzy Association Rule Mining (FARM) which is combined with the selection of complexity metrics using CFS and dataset balancing using SMOTE has sensitivity 85.51% and accuracy 91.63% in detecting software defective modules on NASA MDP dataset.