PREDICTION OF STUDENT ACADEMIC PERFORMANCE AND GRADUATION TIME USING LSTM AND GRU

• Main Author: Kurniawati, Gisela
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/65780
• Institution: Institut Teknologi Bandung

High drop out rates and low student performance are unavoidable problems for educational institutions, both national and international. In order to anticipating the continued growth in the number of students drop out, predicting students who are at risk of drop out is important. Consequently, this study presents an automated technique to predict student performance and graduation using student data with separated and combined prediction method. The data is collected from a study program at a university in Bandung. Data preprocessing is done to produce a dataset with common features to predict the two tasks. Long Short-Term Memory (LSTM) and Gate Recurrent Units (GRU) as an outstanding model in handling sequence data was proposed in this study. According to our study, the LSTM and GRU architectures performed well in predicting each task. The performance of each architecture was surpassed each other depending on the corresponding task. On the task at graduation, the F1-score of the best model has reached 94% and on the student performance task, the F1-score is 86%. While in the combined task the resulting F1-score value is 73%. In early prediction, the prediction of graduation time can provide good performance since the second semester with an F1-score of 83%. Meanwhile, in predicting student achievement, the RMSE value has met the rules of thumb since the second semester of 0.426. Overall student performance and graduation prediction are better when using separate methods compared to combined methods. In this study, the use of dropout layers tends to degrade model performance. Furthermore, the preprocessing data and baseline model generated in this study can be replicated for similar tasks at other universities with feature adjustments based on data availability.