Recognizing Developer Activity Based on Joint Modeling of Code and Command Interactions

Recognizing the current activity of a software developer (e.g., debugging, reading, editing) can improve the effectiveness of recommendation systems that aim to reduce the cognitive load of information lookup during software development. Current recommendation systems based on developer activity detection focus primarily on a single dimension of developer behavior, e.g., both observing and recommending IDE commands or source code classes. In addition, the current state of the art techniques require that the number and type of activities exhibited by developers is pre-specified and that labeled interaction data is provided as input. In this article, we propose the use of an approach that eschews these requirements, leveraging an unsupervised statistical model that uses both IDE commands and source code accesses to discern latent developer activities. Our approach outperforms baseline supervised and unsupervised learning techniques on simulation-based evaluation of source code and command recommendations for most of the configurations we examined. We also show that our technique benefits from observing both commands and code accesses when identifying developer activity.