The Development Of KNN Diagnosis Method And Its Application In Practice

Cognitive Diagnosis Assessment(CDA)is a new generation measurement theory in recent years.CDA has laid the foundation for further education teaching.The realization of the CDA often requires cognitive diagnosis models as the foundation.Cognitive Diagnosis Models(CDMs)are using mathematical model or algorithms to analyze the students' observed scores,and thus estimate the Knowledge State(KS)that show the cognitive structure.Predecessors' researches show that the known cognitive diagnosis models have more than 100 species,but almost all of them are designed based on the parameters.Due to the accurate estimation of parameters needs a variety of conditions of guarantee,such as large sample size and specific algorithms such as EM algorithm or MCMC algorithm,etc.,these conditions are greatly limited the CDA's application in practice.Non-parametric diagnosis methods,however,because do not require the large sample size as the guarantee of accurate estimation,and it show some character can be more easily applied in practice such like easy to understand and operate.Therefore,in recent years,there have been many researches foucus on non-parametric cognitive diagnostic methods.Due to the machine learning algorithm has the advantages of classification efficiency,so in recent years,there are many researchers try to blend in machine learning algorithms in the CDA.many Non-parametric diagnosis method are based on the development and application of machine learning algorithm,such as both K-means diagnosis method and Grade Response Cluster Diagnosis Method(GRCDM)are based on K-means algorithm,the PNN diagnostic method is based on Probabilistic Neural Network.However,almost all of the cognitive diagnosis methods that combined with machine learning algorithms are just applied in the situation of 0-1 score,they do not have the character of dynamic and they are also not simple enough in the meantime.It is necessary to find more simple and more efficient machine learning algorithm applied to the CDA,so as to reach the purpose that open up broader prospects for cognitive diagnosis.K-Nearest Neighbour(KNN)algorithm is called the most simple of a kind of classification algorithm in machine learning algorithms.As a result of KNN directly use of the known categories of data as a training set.it does not involve the use of parameter estimation,so KNN is also a kind of non-parametric algorithm.This paper try to apply KNN algorithm in CD A,which develop a new diagnosis method named KNN diagnosis method that is suitable for polytomous scoring.This paper designed three simulation researchs and empirical research to validate the effectiveness and stability of KNN diagnosis method.The first research discussed how KNN algorithm applied in CDA to become KNN diagnosis method and the influence of sample size on KNN diagnosis method.The second research discussed the accuracy of KNN diagnosis method,and compared it to other non-parametric diagnosis methods;The third research discussed the stability of KNN diagnosis method,which is to compare the Pattern Match Ration(PMR)reduction of KNN diagnosis method and other methods in the case of errors of Q matrix.And in the forth study,a group of empirical data were analyzed by KNN diagnosis method,and compared with MDD-R method.These results are as follow:(1)The study found that KNN algorithm can well combine with the CDA,and the KNN diagnosis method can be applied to score dichotomous response and polytomous response,while KNN diagnosis method is independent of the sample size.;(2)In general,in the case of the correct setting of Q matrix,KNN diagnosis method presents a higher PMR than other non-parametric methods.Among them,the distribution of the subjects will not affect KNN diagnosis method.(3)The increase of the number of attributes and the decrease of the degree of closeness will lead to the decrease of its PMR,and the trend of PMR change in the three other diagnosis methods showed roughly the same,while KNN diagnosis method shows a better reliability in these variables.(4)In the case of errors of Q matrix,KNN shows a more sensitive characteristic of the setting of Q matrix,which means KNN diagnostic method is greatly influenced by the setting of Q matrix;(5)The types of Q matrix error and attribute hierarchy have an impact on the PMR:In the linear and convergence structure,redundant attributes lead to the decline of the minimum,while the other three kinds of structure,attribute redundancy caused the PMR decline is larger.KNN,PNN and MDD-R showed the same trend,while GRCDM is slightly different in linear and convergent structure.(6)KNN diagnostic method has a good empirical reliability and validity,and the result is consistent with the practical situation.Through the above results,the following conclusions can be drawn:(1)KNN diagnosis method has a good classification accuracy,and it is not easy to sample size and the influence of the distribution morphology,so that KNN diagnosis method has a broader application prospect in practice.Under different variables,KNN diagnosis method shows a high accuracy rate,which fully reflects the characteristics of the classification efficiency of machine learning algorithm.(2)KNN diagnosis method is more sensitive to the setting of Q matrix,it is less to be affected by the error of Q matrix as the closer of hierarchical structure.Future research can try to apply it to Q matrix error checking.(3)In the empirical study,KNN diagnosis method shows a good,reliability and validity of diagnostic results,and shows consistency with actual situation.These results show KNN diagnosis method is also applicable in the empirical research.