Concept Graph Learning
A recent WSDM paper by Yang et. al explores a novel machine learning technique called Concept Graph Learning, which can be used to predict prerequisite relationships between educational courses. The technique uses existing graphs of prerequisite dependencies between university courses (e.g., the Cornell CS flowchart) to learn a graph of prerequisite dependencies between “universal concepts”, which are more granular and can be shared across courses offered by different universities. In the input data, a graph consists of nodes that correspond to courses offered by a single university, and directed edges that point from prerequisite courses to postrequisite courses. The model uses multiple input graphs from different universities to build a concept graph that consists of nodes that correspond to universal concepts, and directed edges that point from prerequisite concepts to postrequisite concepts. The model then uses this concept graph to predict unexisting prerequisite edges between courses that have been annotated with concept tags.
Using directed graphs to represent relationships between courses, between courses and concepts, and between concepts, the authors were able to build a model that can be used for a variety of applications, such as algorithmically constructing dependency graphs that link MOOC courses created by different institutions.