Predicting future learning better using quantitative analysis of moment-by-moment learning

In recent years, student modeling has been extended from predicting future student performance on the skills being learned in a tutor to predicting a student’s preparation for future learning (PFL). These methods have predicted PFL from a combination of features of students’ behaviors related to meta-cognition. However, these models have achieved only moderately better performance at predicting PFL than traditional methods for latent knowledge estimation, such as Bayesian Knowledge Tracing. We propose an alternate paradigm for predicting PFL, using quantitative aspects of the moment-by-moment learning graph. This graph represents individual students’ learning over time and is developed using a knowledge-estimation model which infers the degree of learning that occurs at specific moments rather than the student's knowledge state at those moments. As such, we analyze learning trajectories in a fine-grained fashion. This new paradigm achieves substantially better student-level cross-validated prediction of student’s PFL than previous approaches. Particularly, we find that learning which is spread out over time, with multiple instances of significant improvement occurring with substantial gaps between them, is associated with more robust learning than either very steady learning or learning characterized by a single “eureka” moment or a single period of rapid improvement.