A Double-layer Stacked Gate Recurrent Unit with Self-Attention Residual Model for Knowledge Tracing
Liang Qian, Xiaoyu Luo · 2024
In recent years, the rapid development of online tutoring systems has significantly increased research interest n the domains of knowledge tracing and deep learning. Knowledge tracing aims to assess learners' understanding of knowledge points by modeling changes in cognitive states over time, discovering latent learning patterns within learners' activities, and providing personalized guidance, thereby facilitating the application of artificial intelligence in educational paradigms. Deep learning, known for its superior feature extraction capabilities, has shown great potential in enhancing the performance of knowledge tracing models, attracting widespread attention. Despite the advancements, current neural network-based knowledge tracing models exhibit potential for further enhancements in efficiency and efficacy. This paper introduces a deep knowledge tracing model that integrates double-layer stacked Gated Recurrent Units (GRU) and self-attention mechanisms. This model opts for a double-layer stacked GRU over traditional Long Short-Term Memory (LSTM) networks to address the overfitting challenges posed by the latter's extensive parameter set, and integrates residual connections to streamline the training process. Additionally, it combines self-attention mechanisms and dilated convolution techniques to extract relational and local features of the learning sequences. Furthermore, the model introduces a StepLR learning rate decay strategy within both Deep Knowledge Tracing (DKT) and Double-layer Stacked Gate Recurrent Unit with Self-Attention Residual Model(SA-DGRU-R), allowing for dynamic adjustment of the learning rate to refine the training dynamics. Experimental results on the Assist2009 and Assist2015 datasets, evaluated using AUC and F1-scores, demonstrate the model's superior performance over other similar recurrent neural network and attention mechanism models.