Adversarial Multi-task Learning-based Bug Fixing Time and Severity Prediction

Abstract

Bug reports contain important information for software quality assurance. Conventionally, engineers complete bug report analysis tasks, which are extremely burdensome. Recently, researchers and companies have been working towards the automation of bug report analysis. Many machine learning and deep learning models are utilized for triage and prediction of bug report attributes such as bug fixing time and bug severity based on the description and comment text in bug reports. However, due to the rapid growth of data size in bug reporting systems, the prediction accuracy in single-task machine learning models is neither efficient nor effective. Multi-task learning (MTL) is a transfer learning scheme, which can train multiple related tasks together, reducing the training time while improving the overall performance. In our study, we utilize adversarial multi-task learning, which addresses the problem of contaminated shared feature space in common MTL models towards a purer shared feature space. Our adversarial convolutional neural network model (ADV-CNN) improved the validation accuracy of the bug fixing time prediction from 83.67% of a ST-CNN model to 89.25%.