Knowledge tracing is a family of models that estimate, for each learner and each knowledge component (concept, skill, or topic), the probability that the learner has mastered it based on their interaction history — primarily quiz responses and assessment outcomes. The foundational method is Bayesian Knowledge Tracing (BKT), which models mastery as a hidden binary state that updates after each correct or incorrect response using fixed transition and emission probabilities. Deep Knowledge Tracing (DKT), introduced in 2015, replaces the Bayesian model with an LSTM or transformer network that can learn more complex learning patterns from data. In an adaptive learning system, knowledge tracing provides the mastery probability that drives content routing: if the model estimates mastery is above threshold, the learner advances; below threshold, a remedial item is served. Knowledge tracing requires a well-defined knowledge component map — the curriculum must be broken into granular, assessable units — and a sufficient number of practice interactions per unit to make meaningful probability estimates. A practical limitation is data sparsity: many learners attempt each concept only a few times, giving the model little signal. Knowledge tracing is closely related to the recommendation engine, which uses mastery estimates as one signal alongside engagement, time-on-task, and preference data to select the next learning activity.
