Deep Attention Knowledge Tracking Incorporating Multiple Features and TCN-Transformer
Ting Lv, Luqiang Xu · 2024
In contemporary innovative education research, a focal point revolves around applying knowledge tracking to forecast students' prospective academic performance and provide individualized guidance based on their knowledge status. Numerous extant knowledge tracking approaches solely model learning outcomes, neglecting to address diverse learning behaviors comprehensively and thereby underexploiting the plethora of available learning features. Moreover, most current knowledge tracking methodologies predominantly address long-distance dependency issues while overlooking the incorporation of local and global information. In light of these limitations observed in existing knowledge tracking methods, we propose a novel approach—Deep Attention Knowledge Tracking incorporating Multiple Features and TCN-Transformer (MTDAKT). Commencing with a post-hoc perspective, we consider individual student differences. The random forest feature selection method is then employed to extract rich learning features, accompanied by introducing a new feature crossover method. Subsequently, deep attention is applied to augment the model's learning capabilities, effectively capturing the nuanced relationship between practice and response complexities. Finally, coupled with the TCN-Transformer framework, our model integrates global and local information, yielding two distinct knowledge states. Empirical results indicate a substantial improvement, with a 12% increase in AUC and a 13% increase in ACC compared to the traditional knowledge tracking model, demonstrating superior predictive performance and fitting capabilities.