In computer science, introductory programming course is one of the very first courses taken. It sets the base for more advanced courses as programming ability is usually assumed there. Finding the students that are likely to fail the course allows early intervention and more focused help for them. This can potentially lower the risk of dropping out in later studies, because of the lack of fundamental skills. One measure for programming ability is the outcome of a course and the prediction of these outcomes is the focus also in this thesis.
In educational context, differences between courses set huge challenges for traditional machine learning methods as they assume identical distribution in all data. Data collected from different courses can have very different distributions as there are many factors that can change even between consecutive courses such as grading, contents, and platform. To address this challenge transfer learning methods can be used to as they make no such assumption about the distribution. In this thesis, one specific transfer learning algorithm, TrAdaBoost, is evaluated against selection of traditional machine learning algorithms. Methods are evaluated using real-life data from two different introductory programming courses, where contents, participants and grading differ. Main focus is to see how these methods perform in the first weeks of the course that are educationally the most critical moments.
