Design and Deployment of a Better Course Search Tool: Inferring Latent Keywords from Enrollment Networks

Liberal arts universities possess a vast catalog of courses from which students can choose. The common approach to surfacing these courses has been through traditional keyword matching information retrieval. The course catalog descriptions used to match on may, however, be overly brief and omit important topics covered in the course. Furthermore, even if the description is verbose, novice students may use search terms that do not match relevant courses, due to their catalog descriptions being written in the specialized language of a discipline outside of their own. In this work, we design and user test an approach intended to help mitigate these issues by augmenting course catalog descriptions with topic keywords inferred to be relevant to the course by analyzing the information conveyed by student co-enrollment networks. We tune a neural course embedding model based on enrollment sequences, then regress the embedding to a bag-of-words representation of course descriptions. Using this technique, we are able to infer keywords, in a system deployed for a user study, that students (N = 75) rated as more relevant than a word drawn at random from a course’s description.