Modeling Individualization in a Bayesian Networks Implementation of Knowledge Tracing

Individualized Bayesian Knowledge Tracing Models

“…Pardos and Heffernan [5] individualized the initial probability of mastery p(L 0 ) k by assigning according to a set of heuristics: randomly, by selecting from two pre-set values based on first student response correctness, by using overall percent correct. The 'prior per-student' models fit better than traditional BKT on a significant fraction of the problem sets authors considered.…”

“…Although the [potential] benefits of individualized BKT models are visible, the results are unclear about the ideal configuration of student-specific parameters (4 per student [1], 1 heuristic value per student [5], 4 per student [8]), are limited in the evidence for improved mode prediction and are hard to operationalize for the purpose of implementing in an ITS. The original work on BKT [1] pointed out that operationalization of the discussed individualized BKT model could be problematic.…”

“…Starting with the original publication on BKT [1] and including more recent works (e.g. [5]), there exist strong indicators that BKT models (often called individualized BKT models) that somehow account for student variance are superior to the standard BKT model.…”

“…[1], [5]), and [8]) describes quite different approaches to defining and learning student-specific parameters as well as report radically different performance measures. In this paper, we approach the problem of introducing student-specific parameters in a more systematic manner.…”

“…In the standard BKT implementation, there are only skill-specific parameters. However, a large body of research strongly suggests that studentspecific variability in the data, when accounted for, could enhance model accuracy [5,6,8]. In this work, we revisit the problem of introducing student-specific parameters into BKT on a larger scale.…”