Multimodal Machine Learning Model for Automated Bug Report Classification and Root Cause Identification

Abstract

Bug reports contain text, stack traces, source-code references, execution logs, and project metadata, but these sources are often reviewed separately during software maintenance. This study presents a multimodal machine learning model for automated bug report classification and root cause identification. The model processes bug summaries and descriptions, stack traces, source-code context, execution logs, and metadata through separate encoders. An attentionbased fusion layer combines the learned representations and sends them to two prediction branches for bug class and root cause ranking. The model was evaluated using macro F1-score and top-3 root cause accuracy across six modality combinations. The full multimodal configuration achieved the best results, reaching 92.6% macro F1-score and 94.1% top-3 root cause accuracy. The findings show that combining descriptive and technical evidence significantly improves classification and diagnosis. The proposed model can reduce manual triage effort, speed up fault investigation, and support more accurate maintenance decisions in complex software projects.