SATCN: An Improved Temporal Convolutional Neural Network with Self Attention Mechanism for Knowledge Tracing

With the rapid expansion of E-education, knowledge tracing (KT) has become a fundamental mission which traces the formation of learners’ knowledge states and predicts their performance in future learnng activates. Knowledge states of each learner are simulated by estimating their behavior in historical learning activities. There are often numerous questions in online education systems while researches in the past fails to involve massive data together with negative historical data problems, which is mainly limited by data sparsity issues and models. From the model perspective, previous models can hardly capture the long-term dependency of learner historical exercises, and model the individual learning behavior in a consistent manner is also hard to accomplish. Therefore, in this paper, we develop an Improved Temporal Convolutional Neural Network with Self Attention Mechanism for Knowledge Tracing (SATCN). It can take the historical exercises of each learner as input and model the individual learning in a consistent manner that means it can realize personalized knowledge tracking prediction without extra manipulations. Moreover, with the self attention mechanism our model can adjust weights adaptively, thus to intelligently weaken the influence of those negative historical data, and highlight those historical data that have greater impact on the prediction results. We also take attempt count and answer time two features into account, considering proficiency and forgetting of the learners to enrich the input features. Empirical experiments on three widely used real-world public datasets clearly demonstrate that our framework outperforms the presented state-of-the-art models.