A university map of course knowledge

Area of Science:

• Data Science
• Knowledge Representation
• Computational Social Science

Background:

• Digitization generates vast behavioral data, necessitating advanced methods for understanding.
• Natural language processing has shown success in extracting structure from text data.
• Existing methods are insufficient for the full spectrum of big data.

Purpose of the Study:

• To demonstrate that behavioral data can inform latent semantic structure.
• To extract domain knowledge from this behaviorally-informed representation.
• To compare the fidelity of this new representation against traditional course catalog descriptions.

Main Methods:

• Utilized course enrollment histories of 124,000 students to learn vector representations of courses.
• Employed visualization and a novel mapping technique to interpret the learned structure.
• Used semantic interpolation via regression to translate course vectors to catalog descriptions.

Main Results:

• Recovered 88% of course attribute information and 40% of known course relationships from enrollment data.
• Learned representations achieved higher semantic fidelity than course catalog descriptions.
• Identified nuanced content differences and accurately described departments lacking catalog information.

Conclusions:

• Student enrollment data provides a powerful source for learning course semantics and relationships.
• Behavioral data analysis offers a more nuanced understanding than static descriptions.
• This methodology has significant implications for data science and knowledge discovery.