A Comparative Study of Ensemble Learning Techniques for Report-Based Bug Localization

Report-based software bug localization is a task in software maintenance, aimed at identifying the source files responsible for reported bugs. Automating this process can greatly reduce the manual effort required, decrease errors, and save time in the debugging process. This paper assesses the performance of several ensemble learning methods for bug localization, utilizing a diverse set of base learners, including K-Nearest Neighbors, Decision Tree, Multi-Layer Perceptron, Naïve Bayes, Logistic Regression, and TabNet, and implements ensemble techniques such as bagging, boosting, and Random Forest, comparing the effectiveness of individual base learners against various ensemble methods, as well as the performance of different ensemble techniques applied to the same base learners. To mitigate data leakage, we identify and articulate the requirements for train-test splitting specific to this task and propose a grouped rolling-window cross-validation strategy, thereby establishing a more robust evaluation framework. Experimental results indicate that ensemble techniques improve the performance of various base models, with bagging consistently outperforming other methods, while AdaBoost occasionally underperforms. Random Forest stands out as the top-performing ensemble when using decision trees as base learners. Among the implemented models, Bagging-TabNet demonstrates the best performance, though it slightly lags behind state-of-the-art models.

Keywords:

Bug Localization , Ensemble Learning , Information Retrieval , Cross-Validation