On the value of learning from defect dense components for software defect prediction

• Authors: Hongyu Zhang;Adam Nelson;Tim Menzies
• Venue: Proceedings of the 6th International Conference on Predictive Models in Software Engineering
• Year: 2010

BACKGROUND: Defect predictors learned from static code measures can isolate code modules with a higher than usual probability of defects. AIMS: To improve those learners by focusing on the defect-rich portions of the training sets. METHOD: Defect data CM1, KC1, MC1, PC1, PC3 was separated into components. A subset of the projects (selected at random) were set aside for testing. Training sets were generated for a NaiveBayes classifier in two ways. In sample the dense treatment, the components with higher than the median number of defective modules were used for training. In the standard treatment, modules from any component were used for training. Both samples were run against the test set and evaluated using recall, probability of false alarm, and precision. In addition, under sampling and over sampling was performed on the defect data. Each method was repeated in a 10-by-10 cross-validation experiment. RESULTS: Prediction models learned from defect dense components out-performed standard method, under sampling, as well as over sampling. In statistical rankings based on recall, probability of false alarm, and precision, models learned from dense components won 4--5 times more often than any other method, and also lost the least amount of times. CONCLUSIONS: Given training data where most of the defects exist in small numbers of components, better defect predictors can be trained from the defect dense components.