Application of Sparse Attention Mechanism and Bidirectional GRU Neural Networks in Knowledge Tracing

To overcome the limitations of traditional deep knowledge tracing models, particularly in capturing complex exercise relationships and handling long data sequences, we introduce SKT-BiGRU. This optimized model combines bidirectional Gated Recurrent Units (BiGRU), sparse attention, and RMSNorm. It first translates a learner's historical interactions into real-valued vectors for more accurate trajectory modeling. The sparse attention mechanism boosts computational efficiency and interpretability, allowing focus on critical data points for prediction. RMSNorm normalizes the neural network's outputs, improving training stability and convergence speed. The model effectively captures intricate exercise dependencies through self-attention and sparse attention. These hidden states and attention weights are combined with the bidirectional GRU output to predict a learner's performance on upcoming questions.Experimental tests carried out on two public datasets have confirmed that SKT-BiGRU, in comparison to the baseline model, exhibits an improvement in its AUC (Area Under the Curve) metrics ranging from 3.5% to 6.1%, and an increase in ACC (Accuracy) from 0.4% to 3.2%. These results demonstrate its effectiveness in augmenting deep knowledge tracing.