ScholarWorks@중앙대학교

초록

With the recent advancements in artificial intelligence (AI) and the improved performance of mobile devices, Knowledge Tracing (KT) models have gained significant attention in the education area, where they play a crucial role in tracking students' learning progress and providing personalized learning experiences. This study aims to compare the performance of KT models on a Raspberry Pi and a GPU server while examining the feasibility of model light-weighting using the Mixed Precision technique in low-power environments. Specifically, we performed fine-tuning and inference of pretrained models on both low-power devices and high-performance servers. The experimental results showed minimal performance differences in terms of AUC between the two environments, while accuracy (ACC) varied across models. Additionally, Attention-based models experienced significantly longer inference times when AMP was applied on low-power devices, indicating that complex computations require more resources on such devices. Based on these findings, applying AMP to the DKT+ model was identified as the most suitable option for low-power environments. © 2025 IEEE.

키워드