Pull together: Option-weighting-enhanced mixture-of-experts knowledge tracing

• Published in: Expert systems with applications 2024-08, Vol.248, p.123419, Article 123419
• Main Authors: Huang, Tao, Ou, Xinjia, Yang, Huali, Hu, Shengze, Geng, Jing, Xu, Zhuoran, Yang, Zongkai
• Format: Article
• Language: English
• Subjects: Knowledge tracing; Multi-task learning; Option tracing; Option weights
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2024.123419

Educators dynamically adjust their teaching strategies by tracing the development of students’ knowledge states. Knowledge Tracing (KT) plays a role similar to that of educators in online teaching. By analyzing past performances, KT identifies learners’ knowledge states and predicts the outcomes of future exercises. However, the existing KT models assume that the learner’s performance is a binary variable (i.e., correct or incorrect) without refining learner performance or differentiating knowledge states. Multiple-choice tests employ distractors that engage learners in different knowledge states, with each distraction implying a specific error. In multiple-choice exercises, we propose an option-weighting-enhanced mixture-of-expert knowledge tracing (WEKT) method that assigns weights to different options based on improved option weighting scoring. The option weights affirm partial knowledge and refine the knowledge state. Building on the multi-task learning strategy, we design a mixture-of-experts framework that simultaneously predicts correctness and options, traces students’ specific errors, and refines students’ performances. The expert structure combines cognitive theory with deep learning technology, taking into consideration the differences between experts and students. Extensive experiments on large-scale datasets indicate that WEKT can refine knowledge states and attain more precise predictions of student performance. •We employ option weights to refine student performance.•We simultaneously predict correctness and option to identify specific errors.•Combining cognitive theory with deep learning techniques for expert structure.•Three experts achieve information exclusivity and sharing.