AUTOMATED SOFTWARE BUG TRIAGING AND DEVELOPER RECOMMENDATION USING CLASSICAL MACHINE LEARNING, DEEP LEARNING AND LLM-BASED APPROACHES

Abstract

Assigning the incoming bug reports to the best suited developers is a vital but time consuming step within software engineering processes. With the growing size of software systems, the number of received reports has grown to an extent when manual handling becomes unrealistic. Unfortunate bug report assignments increase the time of bug resolution, increase developer workload, cause frequent bug tossing and reduce the quality of produced software. Therefore, automating the process via machine learning and natural language processing techniques has become highly rel evant research directions. This dissertation provides two contributions on automated bug triaging problem solved by means of two methodological approaches. First, this paper provides the comparative analysis of six different classification models, i.e., Support Vector Machine, Decision Tree, Recurrent Neural Network, Convolutional Neural Network, Long Short-Term Memory and Bidirectional Encoder Representations from Transformer (BERT) on real-life bug reports from Mozilla Bugzilla. The analysis includes accuracy, precision, recall, F1-score, training time, testing time and inference latency measures. It was found that BERT provides the highest classification accuracy of 89.20% and F1-score of 87.92%, whereas CNN and LSTM provide competitive accuracy with much less computational cost. Secondly, this paper provides a new approach to bug triaging that considers developers’ workload, namely, workload-aware bug assignment framework on real-life data from Eclipse Bugzilla. The performance of three triaging models is tested: a baseline based on Retrieval-Augmented Generator architecture with Sentence-BERT encoding and majority voting, agentic triaging with a developer workload penalty and Sentence-BERT encoding and finetuned BERT-based sequence classifier. Though the best performing fine-tuned BERT model provides the highest accuracy of 89.25%, agentic triaging framework provides the smallest variation of developer workload 25,424.86, which is 31.8% less than the baseline one.