A Method Of Attribute Calibration Based On D~2 Statistics And Selected Training Set

Cognitive diagnostic assessment is actually a pattern recognition and classification. It can be achieved diagnostic assessment of individual cognitive processes, processing skills or knowledge structure, which is an important progress of psychological measurements.Q-matrix is very important in cognitive diagnosis,it expresses the association between item and attribute. Attributes attached to an item are usually calibrated by experts, however,it is time-consuming,cost unbearable,and the calibration may not be quite right,so, looking for an objective method to define the item attribute is imminent.The research team at Columbia University Statistics Department Jingchen Liu et al proposed a data-driven learning method, this method can effectively correct the Q-matrix contains the wrong questions and to some extent restored the true Q- matrix, but this method is very time-consuming. Yu Xiaofeng has improved it, by using D2 statistics to identify unknown attribute vector, the method has a higher success rate in estimating Q- matrix when the number of attributes is 3,4, but his basic item is randomly selected from the matrix Q,this method makes it lack of representation for Q matrix.Numerous studies show that reachability matrix R plays a very important role in raising accuracy of cognitive diagnosis. In view of this, I propose a concept of selected training set on the basis of Yu Xiaofeng, My study compared the training effect which basic items contain 0, 1, 2 reachability matrix R(denoted by 0R, 1R, 2R) as three kinds of training sets.Monte Carlo results show that, "2R" is slightly better than "1R", "1R" is sharply bette r than "0R",And under the condition of 4 attributes,the results of Monte Carlo simulation show that basic items which contain R has a significant advantage, In addtion,the program’s runtime is also greatly reduced.