Mining Detailed Course Transaction Records for Semantic Information

In studies that generate course recommendations based on similarity, the typical enrollment data used for model training consists only of one record per student-course pair. In this study, we explore and quantify the additional signal present in course transaction data, which includes a more granular account of student administrative interactions with a course, such as wait-listing, enrolling, and dropping. We explore whether the additional non-enrollment records and the transaction data's chronological order play a role in providing more signal. We train skip-gram, FastText, and RoBERTa models on transaction data from five years of course taking histories. We find that the models gain moderate improvements from the extra non-enrollment records, while the chronological order of the transaction data improves the performance of RoBERTa only. The generated embeddings can also predict course features (i.e. the department, its usefulness in satisfying requirements, and whether the course is STEM) with high accuracy. Lastly, we discuss future work on the use of transaction data to predict student characteristics and train course recommender models for degree requirements. [For the complete proceedings, see ED630829.]