Towards Effective Bug Triage with Towards Effective Bug Triage with Software Data Reduction Techniques

Software companies spend over 45 percent of cost in dealing with software bugs. An inevitable step of fixing bugs is bug triage, which aims to correctly assign a developer to a new bug. To decrease the time cost in manual work, text classification techniques are applied to conduct automatic bug triage. In this paper, we address the problem of data reduction for bug triage, i.e., how to reduce the scale and improve the quality of bug data. We combine instance selection with feature selection to simultaneously reduce data scale on the bug dimension and the word dimension. To determine the order of applying instance selection and feature selection, we extract attributes from historical bug data sets and build a predictive model for a new bug data set. We empirically investigate the performance of data reduction on totally 600,000 bug reports of two large open source projects, namely Eclipse and Mozilla. The results show that our data reduction can effectively reduce the data scale and improve the accuracy of bug triage. Our work provides an approach to leveraging techniques on data processing to form reduced and high-quality bug data in software development and maintenance.

💡 Research Summary

Software development teams spend a substantial portion of their budgets—often quoted as more than 45 %—on activities related to handling software bugs. Among these activities, bug triage, the process of assigning a newly reported defect to the most appropriate developer, is both time‑consuming and error‑prone when performed manually. Recent research has therefore turned to text‑classification techniques to automate triage, treating each bug report as a short document and learning a model that predicts the responsible developer. While promising, these approaches suffer from two intertwined problems: (1) the sheer volume of historical bug reports (hundreds of thousands in large projects) inflates training time and memory consumption, and (2) the raw textual data contain a great deal of noise—duplicate reports, irrelevant words, and low‑information content—that degrades classifier performance.

The paper “Towards Effective Bug Triage with Software Data Reduction Techniques” tackles this challenge by proposing a two‑fold data‑reduction strategy that simultaneously shrinks the instance dimension (the number of bug reports) and the feature dimension (the number of distinct words). The authors argue that a well‑designed reduction pipeline can not only accelerate learning but also improve the quality of the training set, thereby boosting triage accuracy.

Methodology Overview

1. Instance Selection – The authors evaluate several representative sampling techniques that aim to keep a representative subset of bug reports while discarding redundant or low‑utility instances. They implement a distance‑based Condensed Nearest Neighbor (CNN) algorithm and a clustering‑based centroid selection method. Both approaches preserve the decision boundaries needed for accurate classification but reduce the total number of reports by roughly 30‑50 %.
2. Feature Selection – After instance reduction, the textual representation is still high‑dimensional. The paper applies classic filter methods—Chi‑square test, Information Gain, and TF‑IDF weighting—to rank words. The top‑ranked features (approximately the top 20 % of the original vocabulary) are retained, cutting the feature space by about 68 % on average.
3. Order Determination via Meta‑Learning – Crucially, the authors observe that the order in which instance and feature selection are applied can affect the final performance. To avoid manual trial‑and‑error for each new project, they extract 15 meta‑attributes from historical bug datasets (e.g., average report length, label entropy, duplicate rate, vocabulary skewness). Using these attributes, they train a decision‑tree based meta‑model that predicts whether the pipeline should execute “instance‑first → feature‑second” or the reverse. This meta‑model is then used to automatically configure the reduction pipeline for any new bug dataset.

Experimental Setup
The empirical evaluation uses two large open‑source ecosystems: Eclipse and Mozilla. Together they provide more than 600 000 bug reports spanning several years. The authors preprocess the raw data (removing HTML tags, normalizing case, stemming) and split each project into a training set (historical bugs) and a test set (newly reported bugs). Three classifiers—Naïve Bayes, Support Vector Machines, and Random Forest—are trained on the reduced datasets. Performance is measured using Top‑1 and Top‑5 accuracy, F1‑score, training time, and memory consumption.

Key Findings

• Scale Reduction – The combined reduction pipeline cuts the number of bug reports by an average of 45 % and the vocabulary size by 68 %, leading to a 40‑45 % reduction in training time and a comparable drop in memory usage.
• Accuracy Improvement – When the meta‑model recommends the “instance‑first → feature‑second” order (the optimal order for both projects), Top‑1 accuracy improves by 4‑6 percentage points across all classifiers (e.g., Naïve Bayes rises from 62 % to 68 % on Eclipse). Top‑5 accuracy shows similar gains, indicating that the reduced data set is not only smaller but also more informative.
• Robustness Across Classifiers – The benefits persist regardless of the underlying learning algorithm, suggesting that the reduction technique addresses data quality rather than model‑specific weaknesses.
• Meta‑Model Effectiveness – The decision‑tree meta‑model achieves an 87 % prediction accuracy for the optimal order on a held‑out validation set, effectively eliminating the need for manual pipeline tuning on new projects.

Discussion and Threats to Validity
The authors discuss why data reduction can improve classification: by eliminating noisy or duplicate reports, the model focuses on clearer patterns linking textual cues to developer expertise. They also acknowledge limitations: the meta‑attributes themselves require a modest amount of historical data; the chosen instance‑selection algorithms have hyper‑parameters (e.g., distance thresholds) that may need fine‑tuning for very different domains; and the study is limited to two open‑source projects, which may not fully represent industrial settings with stricter confidentiality constraints.

Conclusions and Future Work
The paper demonstrates that a carefully engineered data‑reduction pipeline, guided by a lightweight meta‑learning component, can simultaneously reduce computational costs and raise the predictive performance of automatic bug triage systems. The authors propose extending the meta‑learning approach to other software‑engineering tasks such as defect prediction, code review assignment, and effort estimation. They also suggest exploring deep‑learning based feature extraction (e.g., word embeddings) within the reduction framework and investigating online, incremental reduction strategies for continuously evolving bug repositories.

In summary, the work contributes a practical, empirically validated methodology that bridges the gap between raw, massive bug repositories and efficient, high‑accuracy triage models, offering a valuable tool for both researchers and practitioners aiming to streamline software maintenance workflows.