A machine-learning approach to modelling student knowledge, introduced by Piech et al. in 2015, that feeds a learner’s full sequence of exercise interactions into a recurrent neural network which maintains a latent state of mastery and predicts the probability of answering the next item correctly, outperforming Bayesian Knowledge Tracing by learning inter-skill structure and forgetting dynamics directly from data at the cost of interpretability.
Semantic Classification
Content
Definition
Deep Knowledge Tracing (DKT) is the application of deep sequence models to the knowledge-tracing problem: estimating, from a learner’s history of attempted exercises, the evolving state of what they know, in order to predict future performance. Introduced by Piech, Bassen, Huang, Ganguli, Sahami, Guibas, and Sohl-Dickstein at NeurIPS 2015, DKT encodes each interaction as a (skill, correctness) pair and feeds the sequence into a Recurrent Neural Network — an LSTM in the original work — whose hidden state serves as a distributed, continuous representation of the student’s knowledge. At each step the network outputs a probability of correct response for every skill, and it is trained end-to-end on next-answer prediction.
DKT’s significance lies in what it removed. Classical Bayesian Knowledge Tracing models each skill as an independent two-state hidden Markov model with hand-specified structure: separate parameters per skill, no transfer between skills, and no forgetting in the standard formulation. DKT discards these assumptions and lets the network discover inter-skill dependencies, prerequisite structure, learning curves, and forgetting dynamics directly from interaction logs. On benchmarks such as ASSISTments the original paper reported substantial AUC gains over BKT, and the learned models can be probed to recover skill-dependency graphs for curriculum design.
The trade-offs are the familiar ones of deep models. The hidden state is not interpretable as per-skill mastery, which complicates open learner models and teacher-facing dashboards; predictions can behave inconsistently (answering correctly can, counterintuitively, lower a predicted mastery — the “reconstruction and waviness” problem addressed by regularised DKT); and later re-analyses showed that with careful feature engineering, simpler logistic-regression baselines close much of the gap. Subsequent architectures — DKVMN’s memory networks, self-attentive models (SAKT, AKT), and transformer-scale systems such as SAINT trained on the EdNet dataset — define the current state of the art in what is now a standard subfield of educational data mining.
Current Landscape
- Deployment: knowledge-tracing models of the DKT lineage drive item selection in large Adaptive Learning platforms (Duolingo, Riiid/Santa, Squirrel AI) and research Intelligent Tutoring System stacks.
- Benchmarks: ASSISTments, EdNet (130M+ interactions), and Eedi competitions anchor evaluation; AUC on next-response prediction remains the standard metric, with growing criticism of its pedagogical relevance.
- Model families: recurrent (DKT, DKT+), memory-augmented (DKVMN), attention-based (SAKT, AKT, SAINT+), and graph-based (GKT) variants; recent work explores LLM-based tracing and interpretable hybrids that map hidden states back to explicit skill estimates.
- Open issues: interpretability for open learner models, fairness across student subpopulations, cold-start for new items and learners, and the gap between predictive accuracy and better pedagogical decisions.