Investigating Programming Performance Predictability from Embedding Vectors of Coding Behaviors

Abstract

Understanding students' coding behaviors is crucial for providing targeted support in programming education. Automatic analysis of coding behaviors using machines can address the limitations of manual monitoring. Previous studies focused on coding behavior representations without considering differences relative to a model answer. We propose embedding vectors that capture these differences, enabling the distinction between simple and complex code solutions. Evaluating these vectors by predicting assignment scores, we achieved over 15% higher accuracy compared to conventional methods. This approach has the potential to enhance teachers' understanding of students' coding behaviors and improve support in programming education.