Abstract
Knowledge Tracing (KT) is crucial in education assessment, which focuses ondepicting students' learning states and assessing students' mastery ofsubjects. With the rise of modern online learning platforms, particularlymassive open online courses (MOOCs), an abundance of interaction data hasgreatly advanced the development of the KT technology. Previous researchcommonly adopts deterministic representation to capture students' knowledgestates, which neglects the uncertainty during student interactions and thusfails to model the true knowledge state in learning process. In light of this,we propose an Uncertainty-Aware Knowledge Tracing model (UKT) which employsstochastic distribution embeddings to represent the uncertainty in studentinteractions, with a Wasserstein self-attention mechanism designed to capturethe transition of state distribution in student learning behaviors.Additionally, we introduce the aleatory uncertainty-aware contrastive learningloss, which strengthens the model's robustness towards different types ofuncertainties. Extensive experiments on six real-world datasets demonstratethat UKT not only significantly surpasses existing deep learning-based modelsin KT prediction, but also shows unique advantages in handling the uncertaintyof student interactions.