AMPNet: Academic Multi-Task Prediction Network for Students' Grade and Graduation

Machine learning based academic performance prediction of students is showing great significance in ensuring the good quality of higher education. However, most methods only focus on predicting academic performance in a single aspect (such as only grade) or for students in a single Major. To address this issue, this paper designs an academic multi-task prediction network model called AMPNet for both grade prediction and graduation prediction of all majors in universities. It consists of three parts: the LSTM part, attention part, and feature fusion part. The LSTM part extracts the time series features of students' historical terms and predicts the grade points of the next term. Then these features are further refined by the terms and attributes attention part. Finally, the refined features are input into the feature fusion part, concatenating the input major delay-graduation embedding, and outputs the prediction results of both grade and graduation status. Comparative experiments show that the proposed AMPNet model provides good performance in grade point prediction, and it significantly outperforms mainstream machine learning prediction models in graduation prediction, and ablation experiments demonstrate the effectiveness of three key modules: major delay-graduation embedding, terms attention, and attributes attention.