An Enhanced Information Retrieval-Based Bug Localization System with Code Coverage, Stack Traces, and Spectrum Information

Abstract

Several strategies such as Vector Space Model (VSM), revised Vector Space Model (rVSM), and integration of additional elements such as stack trace and previously corrected bug report have been utilized to improve the Information Retrieval (IR) based bug localization process. Most of the existing IR-based approaches make use of source code files without filtering, which eventually increases the search space of the technique, thereby slowing down the bug localization process. This study developed an enhanced IR-based bug localization model as a viable solution. Specifically, an enhanced rVSM (e-rVSM) is developed based on the hybridization of code coverage, stack traces, and spectrum information. Combining the stack trace and spectrum information as additional features can enhance the accuracy of the IR-based technique by boosting the bug localization process. Code coverage analysis was conducted to remove irrelevant source files and reduce the search space of the IR technique. Then the filtered source files are preprocessed via tokenization and stemming from selecting relevant features and removing unwanted words. The preprocessed data is further analyzed by finding similarities between the preprocessed bug reports and source code files using the e-rVSM. Finally, scores for each source code and suspected buggy files are ranked in descending order. The performance of the proposed e-rVSM is tested on two open-source projects (Zxing and SWT), and its effectiveness is assessed using TopN rank (where N = 5, 10), Mean Reciprocal Rank (MRR), and Mean Average Precision (MAP). Findings from the experimental results revealed the effectiveness of e-rVSM in bug localization. In particular, e-rVSM recorded a significant Top 5 (80.2%; 65%) and Top 10 (89.1%; 75%) rank values on SWT and Zxing dataset respectively. Also, the proposed e-rVSM had MRR values of 80% and 54% on the SWT dataset and MAP values of 61.22% and 47.23% on the Zxing dataset.