What are the Most Important Features to Individualize: Extending Knowledge Tracing to Individualize What is Important

One traditional way of modeling student knowledge is Corbett and Anderson Knowledge Tracing [2]. Knowledge Tracing model learns a set of four parameters per skill known in the field as 1) Prior (allowing for the modeling of the average chanced that students know a given skills before starting with the system), 2) Learn (different skills could have different learning rates), 3) Guess (some skills might have different guess rates), and 4) Slip (some skill can have different slip rates). Dozens of EDM papers use this technique and most of them fit different KT models for each skill that they track (assuming students are learning different skills). Note that this traditional method requires each skill be fit separately. Wang and Heffernan ([6], inspired by Pardos & Heffernan, 2010) introduced a unified way of doing Knowledge Tracing such that we can learn four individual parameters per student, while at the same time learning four parameters for each skill. Wang and Heffernan argue that doing this allows personalizing the model to model effects such as that some students might have higher learning rates than others, or different abilities to guess the answers. In this work we look to see what the most important ways of doing this individualization are. For instance, is it more important to individualize student prior knowledge or do you get better predictive accuracy if you model that students could have different learning rates.