Toward Better Grade Prediction via A2GP - An Academic Achievement Inspired Predictive Model

Predicting student performance in an academic institution is important for detecting at-risk students and administering early-intervention strategies. We propose a new grade prediction model that considers three factors: temporal dynamics of prior courses across previous semesters, short-term performance consistency, and relative performance against peers. The proposed architecture comprises modules that incorporate the attention mechanism, a new short-term gated long short-term memory network, and a graph convolutional network to address limitations of existing works that fail to consider the above factors jointly. A weighted fusion layer is used to fuse learned representations of the above three modules—course importance, performance consistency, and relative performance. The aggregated representations are then used for grade prediction which, in turn, is used to classify at-risk students. Experiment results using three datasets obtained from over twenty thousand students across seventeen undergraduate courses show that the proposed model achieves low prediction errors and high F1 scores compared to existing models that predict grades and thereafter identifies at-risk students via a pre-defined threshold.