Using Dirichlet priors to improve model parameter plausibility

Student modeling is a widely used approach to make inference about a student's attributes like knowledge, learning, etc. If we wish to use these models to analyze and better understand student learning there are two problems. First, a model's ability to predict student performance is at best weakly related to the accuracy of any one of its parameters. Second, a commonly used student modeling technique, knowledge tracing, suffers from having multiple sets of parameters providing equally good model fits. Furthermore, common methods for estimating parameters, including conjugate gradient descent and expectation maximization, suffer from finding local maxima that are heavily dependent on their starting values. We propose a technique that estimates Dirichlet priors directly from the data, and show that using those priors produces model parameters that provide a more plausible picture of student knowledge. Although plausibility is difficult to quantify, we employed external measures to show the parameter estimates were indeed improved, even if our model did not predict student behavior any more accurately.