ENSEMBLE-EMPOWERED TF-IDF: SCALABLE NLP FOR AUTOMATED ECLIPSE BUG REPORT CLASSIFICATION

Abstract

Software maintenance in large open-source ecosystems depends heavily on fast and accurate bug 
triage. The Eclipse project alone receives tens of thousands of issue reports annually, covering a wide 
range of components and severity levels. Traditionally, human experts perform this triage manually, a 
process that is time-consuming, inconsistent, and increasingly unmanageable as the scale of the 
project grows. Previous attempts to automate this task using single machine learning models like 
Support Vector Machines (SVM) or Logistic Regression have achieved moderate accuracy 
(approximately 70–75%), but these approaches often rely on extensive feature engineering and 
struggle to generalize across evolving bug datasets. This research proposes a scalable, ensemble-based 
framework for automating Eclipse bug classification with high accuracy. The system processes raw 
bug descriptions through a comprehensive preprocessing pipeline—tokenization, stop word removal, 
and lemmatization—and converts the text into numerical representations using Term Frequency
Inverse Document Frequency (TF-IDF). It then trains and evaluates five classifiers on a curated 
Eclipse–Mozilla dataset: SVM, Random Forest Classifier (RFC), Logistic Regression Classifier 
(LRC), Extra Trees Voting (EV) ensemble, and Extreme Gradient Boosting (XGBoost). A user
friendly GUI is integrated into the system, enabling non-experts to upload data, visualize 
preprocessing steps, and select models. With a 70/30 train-test split, the models yield the following 
performance: SVM achieves 74.23% accuracy, 83.03% precision, 73.94% recall, and 75.24% F₁
score; RFC scores 83.51% accuracy, 87.68% precision, 83.40% recall, and 84.09% F₁; LRC records 
70.10% accuracy, 75.06% precision, 70.19% recall, and 71.10% F₁; EV ensemble achieves 89.69% 
accuracy, 91.10% precision, 90.29% recall, and 90.21% F₁; while XGBoost outperforms all others 
with 92.27% accuracy, 92.91% precision, 92.65% recall, and 92.50% F₁-score. These results 
underscore the strength of ensemble methods, particularly XGBoost, in delivering reliable and 
scalable bug classification for large open-source projects.