Methods for Estimating Seniors' Source Code Helpful for Next Challenge Depending on a Series of Programs Written by Each Student in Recent Exercises

In programming exercises, analysis of the source code created by learners is often used to support instructors. However, most existing studies used this analysis for estimating the learner's understanding or identifying learners who need support. We developed methods for automatically estimating seniors' source code helpful for next challenge depending on a series of codes written by each learner in recent exercises. Specifically, we built a machine learning model that outputs the labeled source code (helpful code) by receiving a series of programs each learner wrote. To build this model, we generate abstract syntax trees by compiling source codes and then convert them to token columns and weights. These are then classified into different solutions (i.e., program logic) using clustering and labeled accordingly. Results of evaluation have shown that our methods can estimate helpful source code with high accuracy.