Predicting Student Performance with Control-flow Graph Embeddings

Existing AST-based techniques for embedding student code have been used to help represent student understanding of programming concepts. However, though these embeddings capture syntactic information about student code submissions, they can miss semantic information like code flow through looping structures and conditionals. To address this limitation, we embedded Control Flow Diagrams of student programming submissions, which include semantic information. For our first evaluation of this technique, we verified that the Euclidean distance between embeddings corresponds to human-evaluated code similarity. Then, we evaluated the effectiveness of this technique for student performance modeling by using these vectors to predict student performance at the end of the course. During our evaluation, we found that our CFG-based embedding approach outperforms our baseline methods by a minimum of 7% when predicting student performance.