A Novel Relevance Aggregation Approach for Bug Localization.

Automated bug localization is the process of identifying the probable cause and the source files that are most likely faulty, based on a given bug report. This task can be approached as an information retrieval (IR) problem, where bugs are treated as queries and the software source code repository is treated as a collection of documents. Recent research has focused on narrowing the lexical gap that exists between natural language used to describe bugs and the identifier naming conventions employed by developers. A wide range of features, such as lexical similarity, semantic similarity, and code change history, have been explored to establish the relationship between a bug report and a source file. However, since these features have varying impacts on classification performance, it is critical to evaluate their importance to determine the fundamental structure that enables a bug report and a source file to match more closely. Numerous criteria exist, each offering a distinct perspective on feature applicability, while considering the entire dataset. Focusing solely on one criterion may lead to the exclusion of crucial elements that another criterion may have identified as essential. This study analyzes the importance of features in determining the relevance of a source file to a bug report from multiple perspectives. A fully connected neural network is used to aggregate relevance scores and produce a more accurate final score indicating the likelihood that a source file is the cause of the bug. Experiments were conducted on six popular open source projects to evaluate the proposed model's performance. The results demonstrate that our approach outperforms several existing bug localization models in terms of top-k accuracy, MAP, and MRR.