Establishing a defect prediction model using a combination of product metrics as predictors via Six Sigma methodology

Abstract

Defect prediction is an important aspect of the Product Development Life Cycle. The rationale in knowing predicted number of functional defects earlier on in the lifecycle, rather than to just find as many defects as possible during testing phase is to determine when to stop testing and ensure all the in-phase defects have been found in-phase before a product is delivered to the intended end user. It also ensures that wider test coverage is put in place to discover the predicted defects. This research is aimed to achieve zero known post release defects of the software delivered to the end user by MIMOS Berhad. To achieve the target, the research effort focuses on establishing a test defect prediction model using Design for Six Sigma methodology in a controlled environment where all the factors contributing to the defects of the product is within MIMOS Berhad. It identifies the requirements for the prediction model and how the model can benefit them. It also outlines the possible predictors associated with defect discovery in the testing phase. Analysis of the repeatability and capability of test engineers in finding defects are demonstrated. This research also describes the process of identifying characteristics of data that need to be collected and how to obtain them. Relationship of customer needs with the technical requirements of the proposed model is then clearly analyzed and explained. Finally, the proposed test defect prediction model is demonstrated via multiple regression analysis. This is achieved by incorporating testing metrics and development-related metrics as the predictors. The achievement of the whole research effort is described at the end of this study together with challenges faced and recommendation for future research work.

Software defect prediction using Bayesian networks

“…In defect prediction literature, there are many defect prediction algorithms studied like regression [43] [10] [40], rule induction [40], decision tree approaches like C4.5 [42], case-based reasoning (CBR) [23] [22] [40], artificial neural networks [24] [44] [21] [40], linear discriminant analysis [31], k-nearest neighbour [6], k-star [25], Bayesian networks [12] [35] [46] and support vector machine based classifiers [26] [19] [20] [41]. According to the no free lunch theorem [45], there is no algorithm which is better than other algorithms on all data sets.…”