Empowering multi-feature deep knowledge tracing with learner forgetting behavior

As an important research focus in the field of smart education, knowledge tracing aims to model the behavior sequences of learners and predicting the probability of learners answering the exercise correctly in the next time. This constitutes the core and pivotal aspect of constructing adaptive educational systems. Current research primarily emphasizes the utilization of single interaction features while often overlooking other heterogeneous features generated during the practice of exercises, which potentially impacts changes in the learner's knowledge state. Additionally, learners experience knowledge forgetting during the learning process, and this forgetting behavior can lead to significant deviations between model predictions and actual outcomes. To address the aforementioned issues and enhance the performance of existing deep knowledge tracing models, a Multi-Feature Deep Knowledge Tracing (MuDKT) model is proposed. MuDKT can effectively model the learner's knowledge state and predict the probability of correctly answering the exercise in the next time by mining these diverse learner heterogeneous features and integrates their forgetting behavior. Specifically, the model consists of three components. Firstly, one-hot encoding is used to represent the learner’s and exercises interaction process based on heterogeneous features and obtain the learner interaction embedding. Secondly, simulating the evolution of learners' knowledge state during the learning process based on the hidden states in recurrent neural network (RNN), and update learners' knowledge state in real-time. At the same time, an attention mechanism with attenuation function is introduced to simulate the forgetting behavior in learners' and exercises interaction. Finally, by applying a sigmoid activation function, the hidden state is passed to a fully connected layer to predict the probability of a learner correctly answering the exercise in the next time. Experimental results on three real-world datasets demonstrate that MuDKT outperforms baselines, which can provide a reference for follow-up personalized learning path generation, etc.