The Last Mile: Adapting Software Analytics into Developers' Context

点击翻译按钮获取中文摘要

The software analytics research area has greatly improved what we might call the "science" of software development, such as understanding latent relationships in the artifact space and the likely accuracy of different prediction methods. However, this research area has also been criticized for a lack of practical results, tools, and techniques that might be deemed "actionable'' by developers and managers. For example, the topic of defect prediction ---- where statistical models based on history and/or program structure are employed to predict which files are the likeliest to contain bugs ---- has perhaps the strongest body of research results in this field. While the usual stated motivation for defect prediction might seem sound ---- extra resources may be devoted to source files marked as defect prone ---- in practice, industry has been slow to adopt these approaches, seeing little practical value to the knowledge gained. Furthermore, while much research effort has been spent exploring various complex relationships among artifacts using heavyweight analysis techniques such as machine learning and natural language processing, the real success stories of industrial software analytics ---- such as are evident on GitHub and the web pages of major open source projects ---- have been in creating simple dashboards based on straightforward metrics, such as tracking developer activities on maintenance tasks. Fundamentally, there is a disconnect between what software analytics researchers spend their time exploring and what information needs industrial developers have; this disconnect is well known within the software analytics research community. The goal of this research program is twofold: to improve modelling and understanding of the information needs of developers and managers, and, by example, to build bridges from these models to techniques from the software analytics research community. In so doing, we will help to span the "last mile" between research practice and industrial needs. This work will aid the Canadian software industry to become more effective in their work processes; developers will be better able to exploit their working context in solving their day-to-day problems, since the tools will be more finely tuned to their actual information needs. In turn, this will result in better informed decision-making by developers, less time wasted, and higher quality software systems overall.