ZeroIn: Characterizing the Data Distributions of Commits in Software Repositories

Modern software development is based on a series of rapid incremental changes collaboratively made to large source code repositories by developers with varying experience and expertise levels. The ZeroIn project is aimed at analyzing the metadata of these dynamic phenomena, including the data on repositories, commits, and developers, to rapidly and accurately mark the quality of commits as they arrive at the repositories. In this context, the present article presents a characterization of the software development metadata in terms of distributions of data that best captures the trends in the datasets. Multiple datasets are analyzed for this purpose, including Stack Overflow on developers' features and GitHub data on over 452 million repositories with 16 million commits. This characterization is intended to make it possible to generate multiple synthetic datasets that can be used in training and testing novel machine learning-based solutions to improve the reliability of software even as it evolves. It is also aimed at serving the development process to exploit the latent correlations among many key feature vectors across the aggregate space of repositories and developers. The data characterization of this article is designed to feed into the machine learning components of ZeroIn, including the application of binary classifiers for early flagging of buggy software commits and the development of graph-based learning methods to exploit sparse connectivity among the sets of repositories, commits, and developers.