One of the key benefits that Bayesian Knowledge Tracing (BKT) offers compared to many competing student modelling paradigms is that its parameters are meaningful and interpretable. These parameters have been used to answer basic research questions and identify content in need of iterative improvement (due to, for instance, low learning or high slip rates). However, a core challenge to the interpretation of BKT parameters is that several combinations of BKT parameters can often fit the same data

more »

... comparably well. Even if, as some have argued, BKT is not truly non-identifiable, in practice highly different parameters with comparable goodness are often found using modern BKT fitting packages. These parameter sets can have highly divergent values for guess and slip. Several approaches have been proposed but none of those have yet led to fully stable and trustworthy parameter estimates. In this work, we propose a new iterative method based on contextual guess and slip estimation that converges to stable estimates for skill-level guess and slip parameters. This method alternates between calculating contextual estimates of guess and slip and estimating skill-level parameters, iterating until convergence. Thus, it produces a more stable set of parameters that can be more confidently used in analyzing content efficacy.