Applying Multitopic Analysis of Bug Reports and CNN algorithm to Bug Severity Prediction

Eontae Kim, Geunseok Yang, Inhong Jung

http://doi.org/10.5626/JOK.2023.50.11.954

Bugs are common in software development. Depending on the severity of bugs, they can be classified as major errors and minor errors. In addition, the severity of the bug can be selected by the bug reporter. However, the bug reporter could apply subjective judgment, which can lead to errors in the severity judgment. To resolve this problem, in this study, we predict the bug severity by applying topic-based Severe and Non-Severe extraction with convolutional neural network (CNN) learning. First, by using the properties of the bug report, is the predicting process is divided into Global topic, Product topic, Component topic and Priority topic and the bug reports are extracted from each topic based on Severe and Non-Severe. The Severe and Non-Severe features are extracted from the Global topics, and severity features are extracted from the Product, Component and Priority topics in the same way. The extracted features are combined, put into the CNN algorithm as an input layer, and the model is trained. To evaluate the efficiency of our model, a comparison between the proposed model and the baselines were conducted in the Eclipse, Mozilla, Apache and KDE open-source projects. Our model showed an improved performance. The results showed 97% for Eclipse, 96% for Mozilla, 95% for Apache and 99% for KDE, showing an average performance improvement of about 24.59% compared to the baseline, and a statistically significant difference.