Automated defect prioritization based on defects resolved at various project periods

Abstract

Defect prioritization is mainly a manual and error-prone task in the current state-of-the-practice. We evaluated the effectiveness of an automated approach that employs supervised machine learning. We used two alternative techniques, namely a Naive Bayes classifier and a Long Short-Term Memory model. We performed an industrial case study with a real project from the consumer electronics domain. We compiled more than 15,000 issues collected over 3 years. We could reach an accuracy level up to 79.36% and we had 3 observations. First, Long Short-Term Memory model has a better accuracy when compared with a Naive Bayes classifier. Second, structured features lead to better accuracy compared to textual descriptions. Third, accuracy is not improved by considering increasingly earlier defects as part of the training data. Increasing the size of the training data even decreases the accuracy compared to the results, when we use data only regarding the recently resolved defects.