Machiine Learning and Visualisation Techniques for Inferring Logical Errors in Student Code Submissions

One of the most difficult tasks in computer-aided learning systems for problem-based laboratories, such as in the mathematical and computing sciences, is providing useful feedback to students on the logical errors they have made in reaching an incorrect solution. This paper addresses this problem in the domain of data structures programming laboratories. The question we seek to answer is can we infer the location of logical errors in students' code submissions from automatically generated behavioural tests. We outline techniques for visualizing and clustering submissions using directed graphs. We then examine the use of neural networks for error identification, and show that despite inherent ambiguities in the data, broad error locations can be suggested with 75-90% accuracy.