Deep learning-based knowledge tracing (DLKT) models student knowledge states using deep neural networks, showing better results than Bayesian Knowledge Tracing (BKT) and Performance Factors Analysis (PFA) without requiring human-engineered features.

In Deep Knowledge Tracing (DKT), student interaction sequences are transformed into fixed-length inputs using one-hot encoding and passed to a recurrent neural network (RNN). The hidden state represents the student's latent knowledge state, which is used to compute the probability $P(a)$ of answering subsequent questions correctly. However, because the RNN captures all information within the same vector space, DKT struggles to provide consistent predictions across time or precisely pinpoint which knowledge components (KCs) a student has mastered.

To address this, Dynamic Key-Value Memory Networks (DKVMN) use two memory matrices: a static key memory storing latent concepts, and a dynamic value memory representing the student's evolving knowledge state. While DKVMN outperforms DKT, its vector representations still lack direct interpretability.