Ensemble approach for predicting bug priority level using deep learning algorithms

Abstract

A bug is an unexpected fault or flaw in software systems or devices, frequently brought on by bad programming. The priority of the bug is a crucial factor to consider when addressing the bug and this bug priority prediction is done manually. However, it was a difficult task and the wrong decision might be a reason for huge software failures. The main objective of this study is to propose an ensemble approach to predict the bug priority level of bug reports. 25,000 bug reports from Bugzilla dataset which include the bug descriptions and priority levels are used. After pre-processing the data by tokenization, stemming, stop-words, and lower-casting, study uses feature extracting techniques namely Glove, Word2Vec, Tf-idf, and Doc2Vec. We use a model that primarily uses eight architectures of Convolutional Neural Network (CNN) : AlexNet, LeNet, VGGNet, 1DCNN, ResNet, LZ Net, DenseNet and Siamese. Then the five architectures :ResNet, DenseNet, LZNet, AlexNet, and 1DCNN that had the best accuracy are used in an ensemble method and the final results were taken by the majority values. The performance of the ensemble approach showed 79.18% accuracy. Other individual architectures show accuracies of AlexNet 77.1 %, LZ Net 75.08 %, VGG Net 60.06%, 1DCNN 75.44 %, ResNet 77.34 %, Siamese 39.98 %, DenseNet 77.32 % and LeNet 48.58 %. It was discovered that the proposed ensemble model performed better than the individual algorithms. Finally, when a new bug is identified it can be added to this proposed model, and the model will then determine the bug's priority level.