Enhanced Student-graph Representation for At-risk Student Detection

Predicting examination grades is essential to facilitate early interventions and to enhance student retention rates in an academic institution. We propose a predictive model based solely on historical academic performance made available before the beginning of each semester. The proposed model employs singular value decomposition to distill the underlying student-course graph structure, resulting in a student representation vector that holistically captures a student’s academic ability across courses in relation to their cohort. This representation vector is then fused with the student’s historical academic records for grade prediction. Data for training the proposed prediction model was sourced from approximately five thousand Electrical and Electronic Engineering students across seventeen core courses, including Circuit Analysis and Analog Electronics taught in the sophomore year. Students identified as at risk of failing a course at the beginning of each semester may be offered targeted (academic) support such as peer tutoring programs.