A Four-Stage dynamic approach for Software Defect Prediction

Defects are a major constraint in the software development process. Defects can cause a lot of trouble if not found in earlier stages of software development. They can result in cost overruns and can even lead to a complete system failure. Many supervised machine learning algorithms have been used over the past few decades for software defect prediction. However, the huge variety of software projects causes a problem of varying performance. No machine learning algorithm can deliver the best results consistently for all datasets. Also, a given dataset can have a variety of tuples having different similarities and dissimilarities. Therefore, there is a need to address these problems while predicting defects in any project. This paper introduces an approach for defect prediction which takes into consideration the variety of software projects available. This approach is based on dividing the training dataset into multiple subsets and applying multiple supervised learning algorithms on each such subset to determine the best algorithms for each such subset and then using this information to predict software defect in future projects by making use of similarities in datasets. The given dataset is divided into four parts based on whether the results of applying any random machine learning algorithm on that dataset were true positive, true negative, false positive, or false negative. Experimental results on the PROMISE data repository show an improvement in performance over the existing machine learning algorithms.