Analysis and Improvement of Nature-Based Bug Report Classification Using Ensemble Learning
Abstract
The process of classifying bug reports serves an essential function in software maintenance because it helps organizations determine which issues require immediate attention and which resources should be assigned to them. Existing methods for ensemble learning still struggle due to limited feature sets and treat all classifiers the same way, missing the diverse information that bug reports contain. The researchers introduce HAN BugNet, an innovative framework that uses hierarchical attention to address current problems through three major innovations. First, a hierarchical multi-field fusion mechanism that processes summary, description, stack traces, and comments, learning both token-level and field-level importance via attention. Second, an adaptive weighted ensemble with meta-learning is proposed to learn optimal combinations of classifiers for each bug category, moving beyond simple voting schemes. The domain-aware contextual augmentation strategy generates text variations that preserve essential technical terms and yield realistic results for addressing class imbalance. The evaluation of HAN BugNet on a benchmark dataset of bug reports achieved an accuracy of 99.12%, surpassing all existing ensemble methods and representing the current best performance. The findings reveal that a comprehensive hierarchical method for bug report representation, combined with an adaptive classifier blending technique, results in precise and dependable classification, establishing a new standard for nature-based bug report assessment.