A Dual-Mode Grade Prediction Architecture for Identifying At-Risk Students

Predicting student performance in an academic institution is important for detecting at-risk students and to administer early intervention strategies. In this work, we develop a new architecture that achieves grade prediction based only on grades achieved over past semesters. Our proposed architecture involves two stages—weighted loss function incorporated to the long short-term memory (LSTM) model in the first stage, followed by a short-term gated LSTM in the second. The weighted loss function in the first stage ensures low prediction error by weighing loss associated with the minority class label (in our case the at-risk label). The short-term gated LSTM in the second stage, on the other hand, models short-term variations in academic performance to suppress any residual false alarms. Experiment results using three datasets obtained from over twenty thousand students across seventeen undergraduate courses show that the proposed model achieves a 28.8% improvement in F1 score compared to the LSTM model for at-risk detection. Students identified as at-risk have also been presented and validated by counselors via a dashboard.