Recent works on Intelligent Tutoring Systems have focused on more complicated knowledge domains, which pose challenges in automated assessment of student performance. In particular, while the system can log every user action and keep track of the student's solution state, it is unable to determine the hidden intermediate steps leading to such state or what the student is trying to achieve. In this paper, we show that this information can be acquired through data mining, along with the type, frequency and context of errors that students made. Our technique has been implemented as part of the student model in a tutor that teaches red-black trees. The system was evaluated on three semesters of student data. Analysis of the results shows that the proposed framework of error analysis can help the system in predicting student performance with good accuracy and the instructor in determining difficulties that students encounter, both individually and collectively as a class. [For the full proceedings, see ED593090.]