Imbalanced classes in data mining have more challenges in the educational data mining field. This is because most of the datasets collected from educational records are imbalanced by nature. Some classes dominate others and cause bias predictions. This paper studies the effects of the imbalanced classes on the performance of seven different classifiers, which are J48, Random Forest, k-Nearest Neighbors, Naïve Bayes, Random Tree, SVM, and Linear Regression. Moreover, the effectiveness of the SMOTE technique for handling imbalanced data is evaluated against these classifiers. This will be done through the proposal of an early predictive model that predicts student’s academic performance and recommends their appropriate department in a multi-disciplinary institute. According to our results, the Random Forest technique is the best and has the highest level of accuracy is 94.585%.