Predicting Non-Completion of Programming Exercises using Action Logs and Keystrokes

Computing programming skills is gaining importance but yet it is not an easy subject to master as exemplified in the high attribution rates of computer science courses. It is thus critical that we identify learners who struggle with computer programming at opportune moments to sustain and support their learning. In this study, the keystrokes and actions of learners are captured in real-time in a Java programming tutoring system and used to predict the possibility of non-completion of programming exercises on a granular basis. The results indicate that this can be detected at an adequate accuracy with the proposed feature engineering and machine learning techniques. The use of keystrokes and action logs is significant as they are unobtrusive and easy to set up and thus can easily be propagated for use in any learning environment.