There is an increase in interest and application of Deep Learning Technique to predict student performance and create a computer-aided recommender system to improve student performance in university. However, data on student performance often include categorical variables. How to encode these variables in a deep learning process can be important in the predictive system as Deep Learning only works well with numerical data. The study used a data set of nearly 3000 students in business schools in Vietnam, with academic records and additional surveyed information. Different categorical variables encoding methods were experimented with several deep learning models, in which categorical variables encoding was either embedded or non-embedded in the learning process to find the most effective encoding method. The study experiment shows the learned embedding combined with the student performance prediction deep learning models give good results. The embedded encoding method improves the performance of both Deep Dense and Long Short Term Memory models, with the average accuracy of 86.26%. The study shows the importance of treating categorical variables in a deep learning system for student performance prediction. The embedded categorical variables encoding is recommended to improve the performance and effectiveness of the prediction models.