An Enhanced Deep Knowledge Tracing Model via Multiband Attention and Quantized Question Embedding

Abstract

Knowledge tracing plays a crucial role in effectively representing learners’ understanding and predicting their future learning progress. However, existing deep knowledge tracing methods, reliant on the forgetting model and Rasch model, often fail to account for the varying rates at which learners forget different knowledge concepts and the variations in question embedding covering the same concept. To address these limitations, this paper introduces an enhanced deep knowledge tracing model that combines the transformer network model with two innovative components. The first component is a multiband attention mechanism, which comprehensively summarizes a learner’s past response history across various temporal scales. By computing attention weights using different decay rates, this mechanism adaptively captures both long-term and short-term interactions for different knowledge concepts. The second component utilizes a quantized question embedding module to effectively capture variations among questions addressing the same knowledge concept. This module represents these differences in a rich embedding space, avoiding overparameterization or overfitting issues. The proposed model is evaluated on popular benchmark datasets, demonstrating its superiority over existing knowledge tracing methods in accuracy. This enhancement holds potential for improving personalized learning systems by providing more precise insights into learners’ progress.