Un algorithme d'echantillonnage permettant d'ameliorer la performance predictive d'un modele bayesien dans le cadre d'applications de tests adaptatifs.

摘要

Bayesian graphical models are commonly used to build student models from data. A number of standard algorithms are available to train Bayesian models from student skills assessment data. These models can assess student knowledge and skills from a few observations. They are useful for Computer Adaptive Testing (CAT), for example, where the test items can be administered in order to maximize the information they will provide. In practice, such data often contain missing values and, under some circumstances, missing values far outnumber observed values. However, when collecting data from test results, one can often choose which values will be present or missing by a consequent test design. We study how to optimize the choice of test items for collecting the data that will be used for training a Bayesian CAT model, such as to maximize the predictive performance of the model. We explore the use of a simple heuristic for test item choice based on the level of uncertainty. The uncertainty of an item is derived from its initial probability of success and, thus, from its difficulty.The results show that this choice does affect model performance and that the heuristic can lead to better performance. Although the study's results are more exploratory than conclusive, they suggest interesting research avenues.