Knowledge Tracing Unplugged: From Data Collection to Model Deployment

Knowledge Tracing (KT) plays a pivotal role in Artificial Intelligence in Education (AIED) by modeling and predicting learners’ mastery of skills over time. While AIED Unplugged aims to adapt AI solutions for resource-constrained environments, integrating KT in such scenarios is challenging due to limited digital interaction and the feasibility of exploring advanced algorithms. This paper introduces KT Unplugged, addressing the gap in prior research by exploring and experimenting with creating, refining, and implementing state-of-the-art KT models within resource-limited contexts. The contributions of this paper are threefold. Firstly, we present and perform a procedure for simulating data collection in unplugged contexts, resulting in a dataset and replicable methodology for future field studies. Secondly, an empirical study focused on developing and validating KT models for resource-constrained devices, employing sophisticated (deep learning) and classical (Bayesian) algorithms. This contribution provides empirical evidence on the performance of KT unplugged, including a pre-trained model for numeracy education. Finally, a technical study assesses the deployment of the pre-trained model on disconnected, low-cost mobile devices, demonstrating the technical feasibility of KT Unplugged with acceptable inference times and maintained predictive power. By addressing the challenges of KT integration in unplugged scenarios, this research opens new avenues for personalized learning, adaptive instruction, and targeted interventions in education settings with limited infrastructure.