FDKT: Deep Knowledge Tracking Model integrating Forgetting and Difficulty Factors
Yong Wu, Zhiheng Xia, Lirui Deng, Kai Guo · 2025
Tracking knowledge proficiency of students is a crucial technology in intelligent education. Traditional knowledge tracing models primarily focus on modeling the surface-level features of students response sequences (e.g., answer correctness and temporal information), while insufficiently addressing the deep behavioral patterns implicit in cognitive processes. Although recent studies have incorporated forgetting and difficulty features, existing methods fail to adequately reveal the dynamic coupling relationships between forgetting behaviors and item difficulty, leading to estimation biases in tracking knowledge state changes of students. This paper achieves interactive modeling of forgetting behaviors and item difficulty through feature fusion of forgetting and difficulty characteristics, thereby proposing a Deep Knowledge Tracking Model Integrating Forgetting and Difficulty Factors (FDKT). Specifically: 1) Feature fusion between exercise difficulty vectors and student forgetting vectors through fully connected layers captures high-order interaction characteristics; 2) Incorporation of forgetting behaviors into deep knowledge tracing framework by transforming them into neural network components, with training stability ensured through gradient clipping and normalization techniques; 3) Construction of fully connected networks based on exercise difficulty features to generate dynamically updated weight coefficients, establishing a collaborative adaptation mechanism between knowledge state updates and item difficulty. Comparative experiments conducted on two public datasets (ASSISTments2012 and ASSISTments2017) between FDKT and six baseline models including Deep Knowledge Tracing (DKT) demonstrate that FDKT achieves superior prediction accuracy.