Adaptive meta-knowledge dictionary learning for incremental knowledge tracing

Across intelligent education, knowledge tracing (KT) is a fundamental problem in realizing personalized education. Recently, several approaches using Recurrent Neural Networks (RNNs) view knowledge tracing problems as predicting a student's future performance based on the student's past learning activities. However, deep learning-based models merely predict the students’ future performance, failing to explain the reason why the students answered correctly or incorrectly. These approaches are seldom investigated from a theoretical perspective, preventing a deep understanding of student learning. In this paper, we first define incremental knowledge tracing problem and then propose a dynamic tracing method to track students’ knowledge changes, a new online learning-based model, named Online Meta-knowledge Dictionary Learning (OMDL). OMDL assumes that student knowledge increment is processed with stationary independent increments and derives the objective function of knowledge tracing from the perspective of probability. More precisely, this method has three innovations: propose an online learning algorithm suitable for hidden Markov data; update the meta-knowledge dictionary and students’ knowledge representation online without collecting all data; and trace student knowledge increment explicitly. The empirical study was conducted on four education public datasets, with student prediction accuracy used as the metric for assessing algorithm effectiveness. Experimental results indicate that the OMDL algorithm outperforms existing knowledge tracing methods, achieving a 20% increase in prediction accuracy across all datasets.