Abstract:
Maintaining software products is a continuous process where feature improvements, code
refactoring, and bug fixing must be done regularly to ensure the quality of the product. Bug
triaging, the process of identifying the most suitable developer to resolve a bug, is a manual
process that is error-prone and resource-intensive. While automated bug triaging systems exist,
these systems often struggle with inaccurate predictions and performance limitations due to
unstructured and inconsistent textual data.
The proposed model aims to significantly improve the predictive accuracy of automated bug
triaging systems using an advanced hybrid deep learning architecture. Initially, the dataset is
preprocessed and tokenized using the transformer-based BERT model, leveraging its
contextual understanding of raw textual data. To further improve predictive performance, this
study integrates Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term
Memory (Bi-LSTM) networks, combining CNN’s feature extraction capabilities with BiLSTM’s bidirectional contextual understanding.
The model is evaluated using open-source datasets and benchmarked against multiple deep
learning and machine learning models. The proposed hybrid model achieved an accuracy of
86% and an F1-score of 85%, highlighting its strong ability to predict the most suitable
developer. Comparisons with other models indicate that integrating Bi-LSTM improves
accuracy, outperforming the BERT-CNN, which gained 84% accuracy. Even though the
XGBoost model achieved the highest accuracy of 96%, further testing revealed that it lacks
contextual understanding, leading to misclassifications. Those findings indicate the proposed
model’s ability to make accurate predictions, making it a more reliable solution for bug
triaging.
