Enhanced deep knowledge tracing via synthetic embeddings

Knowledge tracing, a mission to estimate a learner’s knowledge level over time, can be simulated by a machine. With the booming of artificial intelligence, more powerful machine learning-driven models have emerged. Plenty of them show satisfactory results. However, inner dynamic relationships among learning components, proper hardness design, and comprehensive consideration of structures remain to be tackled. In this paper, we designed a novel knowledge tracing architecture called synthetic separated self-attentive neural knowledge tracing (SYNSAINT) which mined inner relations among skills with synthesized latent necessary embeddings. It contains an intelligent knowledge system including five synthetic embeddings: skill, hardness, position, elapsed time, and response. Specifically, we mapped questions into a graph and harnessed the power of unsupervised graph representation learning with clustering techniques to obtain the skillcluster embedding. Moreover, we utilized a sub-neural network method to assign a random weight for each item in the hardness measurement. We built deep sequential attentive structures and verified the method on two representative open-source knowledge tracing datasets. To the best of our knowledge, our approach was proven superior to other state-of-the-art knowledge tracing methods. We also implemented abundant ablation studies to demonstrate the effects of our proposed architecture and a case study to depict its dynamic tracing effects.