Data-Driven Learning of Q-Matrix
The full estimate method of the Q-Matrix was referred in this article. The major ideal is computing the distance between the probability response which condition on the possible Q-Matrix and the observed probability of response. Function 3 is the probability of possible response and Function 4 is observed probability of response which is similar with classical difficulty. Function 14 is the distance we want to minimize. The result shows that the recovery rate is great in the large sample size and the specific stop rule of iteration of estimate can improve the recovering in the small sample size. When the part of items needed to be estimate, the estimated item with more attributes will be difficult to be estimate well.
Comment
1. I think the Q’ above the function 10 in the page 553 was typed wrong. Should it be Q’=[1 1; 0 1; 1 0] ?
2. In this method, whether the more difficult item tend to be with more attributes and vice versa? This method strongly relies to the item difficulty. I am worry if there is a unnecessary attribute list in Q-Matrix, the attribute will be misspecific to items under this Data driven method.