Recognizing and understanding user behaviors from screencasts

Abstract

User interacts with computers or mobile devices, leading to user behaviors on screen. In the context of software engineering, analyzing user behavior enables many applications such as intelligent bug fix, code completion and knowledge recommendation for developers. Such technique can be extended to more general knowledge worker environment, in which users have to manipulate devices according to specific guidelines. Existing works rely heavily on software instrumentation to obtain user actions from operation systems, which is hard to deploy and maintain. In addition, considering the security and privacy of some scenarios, non-intrusive is the major requirement to be included in the system.

In this work, we leverage Computer Vision and Natural Language Processing techniques to recognize and understand user behaviors from screencasts, which is a non-intrusive and cross-platform method. We first recognize 10 categories of low level user actions such as mouse moving and type text, then summarize them to higher level abstractions (i.e. line-granularity coding steps). We also try to interpret user interaction with applications by multi-task learning and generate structured language descriptions (i.e. command, widget and location). Finally, unsupervised learning method is introduced for GUI linting problem, which is taken as a case study of user behavior analysis. To train the deep neural networks, we collect diverse video data from YouTube, Twitch and Bugzilla, and manually label them to build the dataset. The experiment results demonstrate the high performance of proposed method, and the user study validate the practical applications of many downstream tasks. Show more