FAST: Feature-Aware Student Knowledge Tracing

Various kinds of e-learning systems, such as Massively Open Online Courses and intelligent tutoring systems, are now producing large amounts of feature-rich data from students solving items at different levels of proficiency over time. To analyze such data, researchers often use Knowledge Tracing [4], a 20-year old method that has become the de-facto standard for inferring student knowledge from performance data. Knowledge Tracing uses Hidden Markov Models (HMM) to estimate the latent cognitive state (student knowledge) from the students’ performance answering items. Since the original Knowledge Tracing formulation does not allow to model general features, a considerable amount of research has focused on ad-hoc modifications to the Knowledge Tracing algorithm to enable modeling a specific feature of interest. This has led to a plethora of different Knowledge Tracing reformulations for very specific purposes. For example, Pardos et al. [5] proposed a new model to measure the effect of students’ individual characteristics, Beck et al. [2] modified Knowledge Tracing to assess the effect of help in a tutor system, and Xu and Mostow [7] proposed a new model that allows measuring the effect of subskills. These ad hoc models are successful for their own specific purpose, but they do not generalize to arbitrary features. Other student modeling methods which allow more flexible features have been proposed. For example, Performance Factor Analysis [6] uses logistic regression to model arbitrary features, but unfortunately it does not make inferences of whether the student has learned a skill.