Automatical Graph-based Knowledge Tracing

Abstract

Knowledge tracing (KT) is an essential task in online education, which dynamically assesses students' mastery of concepts by predicting the probability that they correctly answer questions. One of the most effective solutions for knowledge tracing is graph-based methods. They maintain multiple vectors to represent students' mastery of concepts, and use these vectors to predict the probability of students correctly answering questions. To give more accurate predictions, the graph-based methods require concept relations to update these vectors once students answer questions. However, the concept relations usually require manual annotation in a real-world scenario, limiting the application of the graph-based method. In this paper, we proposed a method called Automatical Graph-based Knowledge Tracing (AGKT), which is a graph-based method that updates these vectors without requiring manually annotated concept relations. We evaluate our method on four public datasets and compare it with ten advanced methods. The experiment results demonstrate that AGKT yields superior performance.