A Research Of Multilevel Cognitive Diagnosis Model Based On DINA Model

With the development of machine learning technology,the Intelligent Tutor System can better guide students' personality.The core of the Intelligent Tutor System is the student model,but the current student model summarizes the student knowledge based on the computer adaptive technology of the Item Response Theory,and obtains only a single ability score.This score cannot be accurately applied in student targeted learning.As a new generation of test theory,cognitive diagnosis theory can accurately diagnose students' specific knowledge of each knowledge point compared with Item Response Theory.According to the ability score of each knowledge point,cognitive diagnosis theory provides efficient and Personalized guidance for each student,which promote the comprehensive development of students' moral,intellectual,physical and mental skills.The Ministry of Education and the National Development and Reform Commission have put forward an educational policy that combines disciplines such as Artificial Intelligence,cognitive psychology,and neuroscience to improve the discipline distribution of Artificial Intelligence in our country and promote the application of Artificial Intelligence in education.In the future,cognitive diagnosis research will become an important research direction for the combination of Artificial Intelligence and education.DINA model had widely used in cognitive diagnosis because of its high diagnostic accuracy and strong interpretability.However,the DINA model only divided the potential knowledge status of the subjects only two class,which decreases the accuracy of the model in multi-level scoring data.Therefore,based on the DINA and R-DINA models,this paper proposes T-DINA and M-DINA models suitable for multi-level scoring data,and evaluates the parameter estimates of the two models based on the EM algorithm.The main contributions of this article are:1)Based on the fact that most students just master partial knowledge in teaching environment,this paper optimized the class method for the knowledge state of the DINA model.Adding partial mastery knowledge state,the no full mastered state in DINA model is divided into partial mastery and zero mastery.We use the above three states to create the T-DINA model,which improves the shortcomings of the DINA model only applicable to 0-1 score data.2)Analyzing the advantages and disadvantages of the P-DINA model from the mathematical level and combining the design ideas of the T-DINA model,we divide the number of potential knowledge states according to the number of test attributes,and then construct a new K+1 type M-DINA model.Compared with the P-DINA model,the M-DINA model eliminates the constraint relationship between different potential knowledge states,which can improve the accuracy of the model parameter estimation and thereby improve the diagnostic performance of the model;3)Based on the EM algorithm,the parameter estimation procedures of the T-DINA and M-DINA models had implemented.The ABSE and RMSE of the parameters estimated using this procedure are within 2.5%,indicating that the model's parameter estimation procedure has high accuracy;4)Design a comparative experiment,and the result proves that the proposed model has better diagnostic performance than the DINA model.5)Under the independent attribute relationship,the average MMR difference of the three models is within 1%,and the PMR difference of the T-DINA and M-DINA models is within 1%,which is 4.6% higher than the PMR of the DINA model.Under the hierarchical relationship attribute,the MMR different of the three models is within 1.5%,and the PMR of the M-DINA and T-DINA models is 6.6% and 3.2% higher than DINA model respectively.The results show that the M-DINA model has the best diagnostic performance under the independent attribute relationship and the structural attribute hierarchy relationship.,followed by the T-DINA model,and the worst is the DINA model.