FCP2Vec: Deep Learning-Based Approach to Software Change Prediction by Learning Co-Changing Patterns from Changelogs

Abstract

As software systems evolve, they become more complex and larger, creating challenges in predicting change propagation while maintaining system stability and functionality. Existing studies have explored extracting co-change patterns from changelog data using data-driven methods such as dependency networks; however, these approaches suffer from scalability issues and limited focus on high-level abstraction (package level). This article addresses these research gaps by proposing a file-level change propagation to vector (FCP2Vec) approach. FCP2Vec is a recommendation system designed to aid developers by suggesting files that may undergo change propagation subsequently, based on the file being presently worked on. We carried out a case study utilizing three publicly available datasets: Vuze, Spring Framework, and Elasticsearch. These datasets, which consist of open-source Java-based software development changelogs, were extracted from version control systems. Our technique learns the historical development sequence of transactional software changelog data using a skip-gram method with negative sampling and unsupervised nearest neighbors. We validate our approach by analyzing historical data from the software development changelog for more than ten years. Using multiple metrics, such as the normalized discounted cumulative gain at K (NDCG@K) and the hit ratio at K (HR@K), we achieved an average HR@K of 0.34 at the file level and an average HR@K of 0.49 at the package level across the three datasets. These results confirm the effectiveness of the FCP2Vec method in predicting the next change propagation from historical changelog data, addressing the identified research gap, and show a 21% better accuracy than in the previous study at the package level.

Enhancing Software Co-Change Prediction: Leveraging Hybrid Approaches for Improved Accuracy

“…This model aims to transform source code into numerical vectors that preserve semantic and syntactic characteristics of the code, enabling deep learning models to effectively analyze and classify code for various software engineering tasks such as semantic labelling of code snippets, captioning a block of code, generating code to complete a missing piece of a larger program and defect prediction. In the field of co-change prediction, a promising model, FCP2Vec, inspired of Code2vec is proposed in [5]. FCP2Vec (File-level Change Propagation to vector, aims to represent file names in cochange instances extracted from the changelogs of a software project to vector representations that preserve co-change patterns.…”

“…Our approach builds upon the previous work of Lee al. [4], [5], who treated the co-change prediction problem as a recommendation system problem. In this context, the goal is to recommend the top K elements (source files) that are likely to co-change with the currently edited source file (the query element).…”

“…This capability stems from the model's capacity to suggest file names whose vector representations closely match that of the query file name or the vectors of its sub-tokens. In [5], the authors utilized the K ranking score using unsupervised nearest neighbor (UNN) [36]. They adopted a uniform interface for three different neighbor algorithms: ball tree, KDT tree, and brute force.…”

“…After cloning the repository, we parse it to extract the list of all commits done on the project and filter them, as recommended in [5] by removing commits that involve less than two source files and commits that involve more than 50 source files. After that, we form the co-change instances by extracting file names of the co-changed files from each commit.…”

“…Previous studies that utilized changelogs data in cochange prediction have demonstrated that the context provided by co-change instances, as captured by dependency networks [4] or word embedding techniques [5], offers valuable insights into source files that tend to change together. Word embedding techniques enable the creation of low-dimensional representations of sets of tokens or words (source file names), placing co-occurring tokens closer together in the resulting vector space [5]. However, prior studies overlook an important characteristic induced by naming conventions, such as PascalCase and camelCase, which developers adopt to maintain naming consistency throughout projects.…”