Ensemble Knowledge Tracing: Modeling interactions in learning process. (30th November 2022)

• Record Type: Journal Article
• Authors: Sun, Jianwen; Zou, Rui; Liang, Ruxia; Gao, Lu; Liu, Sannyuya; Li, Qing; Zhang, Kai; Jiang, Lulu
• Is Part Of: Expert systems with applications. Volume 207(2022)
• Publication Date: 2022-11-30
• Subjects: Knowledge Tracing -- Deep neural network -- Learning interactions -- User modeling; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.117680
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Knowledge Tracing (KT) aims to continuously estimate students' evolving knowledge state during their learning process, which has attracted much research attention due to its potential for delivering personalized and optimal experiences to students in intelligent learning systems. The learning process is essentially the pairwise interactions of Students, Concepts, and Questions (S–C, S–Q, C–Q for short). Modeling all these interactions will improve the performance of KT. However, existing KT methods hardly exploit all the interactions in a single model. Specifically, Bayesian Knowledge Tracing (BKT) and most of its variants neglect C–Q; Deep Knowledge Tracing (DKT) and other deep neural network approaches mostly neglect S–Q and C–Q. We propose the Ensemble Knowledge Tracing (EnKT), which models all three types of interactions. The base model of EnKT is a hybrid of BKT and DKT. We also present an ensemble algorithm Recurrent Boosting (RB), which extends AdaBoost to deal with KT sequential data. Inspired by BKT, EnKT represents S–C and S–Q using learning and performance parameters, respectively. Besides, EnKT defines C–Q as the correlation complexity among the concepts involved in a question. Experiments show EnKT significantly outperforms state-of-the-art methods (by up to 6% in AUC in some cases) on four real-world benchmark datasets and illustrate better interpretability by several typical case studies. Highlights: Our model represents the interactions among students, concepts and questions. Our model can help teachers understand the status of students. Our model transforms AdaBoost to deal with sequential data. Our model outperforms the traditional knowledge tracing models.