SHF:Small: Mega-Transfer: On the Value of Learning from 10,000+ Software Projects

点击翻译按钮获取中文摘要

As more and more software and software-development artifacts have accumulated in large repositories, Software Engineering has become a Big Data Science. Software analytics is a subfield of Software Engineering in which the software-project data is analyzed to gain actionable information that helps practitioners track and improve software-development practices. Research in software analytics seeks to develop better models, methods and tools to support, for example, software-defect prediction, using software production and programmer activity data. The first decade of software analytics has generated specific results about specific projects, but it remains difficult to generalize across many projects so that results on one project can be applied to other projects. As in any big-data science, a central question is how much data, and of what quality, is needed to draw valid and useful conclusions that can generalize across multiple projects. A typical research study in software analytics looks at no more than a few dozen projects, which limits the validity and usefulness of results in the field. The project will explore the sufficiency and limits of data for transfer learning across domains. The ultimate goal is to scale software analytics, particularly transfer learning of nine extensively studied software-analytics tasks, to 10,000+ open-source projects and 1,500+ commercial projects. The result is software-quality prediction models, backed by large-scale quantitative studies, that can guide the development and maintenance of software, thus reducing the overall cost of creating and maintaining software.The technical challenges of this work include scalability of the algorithms and the ability to transfer lessons learned between projects. To address these challenges this project will develop innovative transfer-learning methods based on fast linear-time clustering algorithms and fast stream-mining algorithms that use incremental hyper-parameter optimization and clustering. The techniques will find similar projects in instance space (within each cluster) in order to find candidates for transfer. Further, where possible, this project will reduce the transfer cost by pre-studies that perform feature selection on the source and target data. Where it is found that software knowledge is highly localized, then this project will also develop and deploy optimization methods for large scale instance-based learning methods.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.