Analyzing Hybrid Feature Representations for Improved Multiclass Bug Severity Classification

The severity classification of software bugs plays a crucial role in the field of software maintenance, as it enables developers to prioritize issues. The present study investigates the effectiveness of hybrid feature representations in improving multiclass bug severity classification using Mozilla bug reports, which contain five predefined severity levels. This study explores the impact of integrating traditional statistical features—Term Frequency-Inverse Document Frequency (TF-IDF)—with contextual word embeddings, including Word2Vec, FastText, and Bidirectional Encoder Representations from Transformers (BERT), to enhance classification performance. Additionally, an assessment is conducted to determine the influence of feature selection techniques, including no selection (all features), Least Absolute Shrinkage and Selection Operator (LASSO), and Principal Component Analysis (PCA), on model performance and training efficiency. The classification performance is measured using three machine learning models: Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF). The results demonstrate that incorporating word embeddings with TF-IDF consistently improves the performance of LR across all cases, achieving an accuracy range of 65.95%–66.15%, compared to 65.69% with TF-IDF alone. Furthermore, applying LASSO for feature selection has been shown to significantly reduce training time while enhancing the performance of SVM. However, the efficacy of hybrid feature representations was found to be less effective for RF. These findings highlight the benefits of hybrid feature representations and feature selection techniques in text-based multiclass classification tasks. This research provides valuable insights for optimizing bug severity classification and can be extended to other domains requiring effective text classification strategies.