Jiayi Zhang, Ryan S. Baker, Debshila Basu Mallick, Cristina Heffernan, Neil Heffernan — cs.HC 📄 Full text (arXiv)
This paper replicates and extends prior work on the cold-start problem in knowledge tracing — the challenge of making accurate predictions when a student begins practicing a new skill. Using a more recent ASSISTments dataset (FoundationalASSIST), the study evaluates KT model performance across both practice trajectories and four problem types: fill-in-the-blank, multiple-choice select-one, multiple-choice select-all, and order/sort. Results show that KT model performance varies across both dimensions, with deep-learning-based models maintaining advantages during early practice but with context-dependent consistency. The study also serves as a proof of concept for SafeInsights, a privacy-preserving research infrastructure designed to facilitate reproducible educational data mining research. This work extends knowledge-tracing-irt findings by demonstrating that problem type — not just skill — moderates model performance, with implications for student-modeling in adaptive learning systems.
Related Pages
- knowledge-tracing-irt — Knowledge tracing models and IRT for student modeling
- intelligent-tutoring — AI tutoring systems and adaptive instruction
- learning-analytics — Data-driven analysis of learning processes
- ai-literacy — Understanding and evaluating AI tools
- student-modeling — Representing learner knowledge and behavior
- cs-education — Computing education research and pedagogy
- self-regulated-learning — Metacognitive strategies for independent learning
- formative-assessment — Ongoing assessment to inform instruction
- automated-grading — AI-assisted evaluation of student work
- scaffolding — Instructional support that fades with competence
Citations
APA: Jiayi Zhang, Ryan S. Baker, Debshila Basu Mallick, Cristina Heffernan, Neil Heffernan (2026). Reexamining the Cold-Start Problem in Knowledge Tracing Models and Implications for SafeInsights. arXiv:2606.11004. cs.HC.