Educational Data Mining and Learning Analytics in virtual environments can be used to diagnose student performance prob-lems at an early stage. Information that is useful for guiding the decisions of teachers managing academic training, so thatstudents can successfully complete their course. However, student interaction patterns may vary depending on the knowledgedomain. Our aim is to design a framework applicable to online Social Sciences and STEM courses, recommending methods forbuilding accurate early performance prediction models. A large-scale comparative study of the accuracy of multiple classifiersapplied to classify the interaction logs of 32,593 students from 9 Social Sciences and 13 STEM courses is presented. Corroboratingthe results of other works, it was observed that high early performance prediction accuracy was obtained based on nothing otherthan student logs: accuracies of 0.75 in the 10th week, 0.80 in the 20th week, 0.85 in the 30th week and 0.90 in the 40th week.However, accuracy rates were observed to vary significantly, in relation to the classification algorithm and the knowledge do-main (Social Sciences vs. STEM). These predictions are generally less accurate for Social Sciences compared to STEM courses,especially at the beginning of the course, with fewer differences observed in the final weeks. Additionally, this research identifiesinstances of low-accuracy outliers in the prediction of Social Sciences courses over time. These findings highlight the complexchallenges and variations in early performance prediction across different domains in online education.
