Improving Contextual Models of Guessing and Slipping with a Truncated Training Set

A recent innovation in student knowledge modeling is the replacement of static estimates of the probability that a student has guessed or slipped with more contextual estimation of these probabilities [2], significantly improving prediction of future performance in one case. We extend this method by adjusting the training set used to develop the contextual models of guessing and slipping, removing training examples where the prior probability that the student knew the skill was very high or very low. We show that this adjustment significantly improves prediction of future performance, relative to previous methods, within data sets from three different Cognitive Tutors.