MODELLING CURRICULA WITH GNN AND LSTM FOR LINK AND SEQUENCE PREDICTION

Abstract

A mathematical model of a university curriculum is proposed. A prerequisite structure is represented as a directed graph, while course order is represented as a sequence. Courses are represented by embeddings that are obtained from syllabus texts and metadata. These vectors are refined by a two layer graph convolutional network (GCN), which is used for link prediction of missing or potentially incorrect prerequisite relations. Course order is represented by a Long Short Term Memory (LSTM) network, which predicts the next course from a fixed window of previously completed courses. An experiment on two bachelor programmes, which are Information Systems and Electric Power Engineering, is reported for 92 courses and 90 explicit prerequisite links. Acceptable quality is obtained for link prediction and sequence reconstruction. Compact tables of recommended new prerequisites are produced for both programmes, which are suitable for curriculum revision decisions.