Bug Localization from Bug Reports: A Multi-Objective Approach

The evolution of software systems has led to increased complexity, making bug localization a labor-intensive task. Early approaches relied on static analysis and manual debugging, which are inefficient for large codebases. Information retrieval techniques such as VSM, LDA, and IR-based tools like BugLocator and BLUiR attempted to automate this process but faced limitations due to lexical mismatches and semantic gaps. Deep learning models like DNNLOC and BugTranslator introduced semantic understanding, yet struggled with scalability and cross-language applicability. Recently, evolutionary algorithms like NSGA-II and MOEA/D have been employed to optimize candidate ranking, but often focus on single objectives, leading to trade-offs between relevance and candidate set size. The current trend emphasizes hybrid models that leverage multiple sources of information and multi-objective optimization to improve accuracy and efficiency.

Despite advances, existing bug localization methods often fail to balance the trade-off between high relevance and minimal candidate size, especially in large, multi-language projects. Lexical similarity alone cannot capture semantic nuances, leading to false positives or missed bugs. Deep models, while powerful, are computationally expensive and less adaptable across different programming languages. Moreover, many approaches lack the ability to incorporate historical bug fix data effectively, which is crucial for predicting recurring issues. The core challenge is designing a scalable, accurate, and language-agnostic system that can handle noisy reports and diverse codebases, reducing manual effort and improving developer productivity.

This work introduces a multi-objective framework based on SPEA-2 that explicitly models the conflicting goals of maximizing relevance and minimizing candidate classes. It innovatively integrates API documentation matching, historical bug fix data, and lexical features within a unified evolutionary process. The novel tokenization method selectively processes meaningful code elements, reducing noise and improving feature quality. The approach also demonstrates cross-language adaptability by validating on both Java and Kotlin projects. These innovations collectively address the limitations of prior IR and deep learning methods, offering a more balanced, scalable, and effective bug localization solution.

• �� Data collection: Gather bug reports, source code, API docs, and historical fixes from six Java projects and one Kotlin project. • Feature extraction: Compute lexical similarity via tokenization, stop-word removal, stemming, and vectorization; match API documentation keywords; incorporate history-based metrics like recent fixes and modification dates. • Model formulation: Define two conflicting objectives—maximize similarity scores and minimize candidate classes—using a multi-objective model. • Optimization: Apply SPEA-2, MOEA/D, and enhanced NSGA-II with opposition-based learning to explore the search space. • Preprocessing: Use a new tokenization strategy focusing on comments, class names, and function identifiers, filtering irrelevant files. • Parameter tuning: Adjust hyperparameters through cross-validation, ensuring robustness across datasets.

Experiments involve six Java open-source projects and one Kotlin industrial project, with metrics including precision, recall, and Top-K coverage. Baselines include NSGA-II, MOEA/D, and traditional IR methods. The evaluation employs k-fold cross-validation, ablation studies on tokenization impact, and cross-language testing. Hyperparameters are optimized via grid search. Results are statistically validated using significance tests, demonstrating the superiority of the proposed framework in both accuracy and efficiency. The experiments also analyze the effect of different feature combinations and the robustness against noisy bug reports.

The proposed SPEA-2 approach achieved a Top-10 accuracy of 88.5% and Top-20 accuracy of 94% across Java projects, outperforming NSGA-II and MOEA/D by 10-15%. Cross-language testing on Kotlin showed comparable performance, confirming model adaptability. The new tokenization technique contributed over 15% improvement in IR-based methods. The multi-source feature integration effectively balanced relevance and candidate set size, reducing manual effort. The results validate the effectiveness of multi-objective optimization in bug localization, especially in large, diverse codebases.

This framework can be integrated into IDEs and CI/CD pipelines to provide real-time bug suggestions, significantly reducing debugging time. It is suitable for large-scale enterprise systems, open-source projects, and multi-language environments. Developers can leverage it for faster fault diagnosis, while project managers benefit from improved software reliability. Future integration with semantic models could further automate bug triaging, making maintenance more efficient.

The approach's effectiveness diminishes with poorly described bug reports lacking clear keywords or API references. Computational costs remain high for very large repositories, requiring further optimization. Support for emerging languages like Rust or Go is limited, necessitating additional feature engineering. Handling noisy or ambiguous reports remains challenging, and future work should focus on robustness and scalability enhancements.