Research On The Method And Application Of Students' Cognitive Diagnosis Based On Deep Learning

In the online learning platform,in order to effectively carry out personalized learning,there is a core task called cognitive diagnosis(CD).It can be divided into two categories:static and dynamic cognitive diagnosis.Static cognitive diagnosis is an overall study of students' learning data in a certain period of time.Comprehensive analysis of these data can get and only get students'current level of knowledge mastery,and then predict students' performance on unobserved questions,without considering the dynamics of students' learning process.The existing methods only use the index of students and test questions and the records of students' answers as the research data,so they are only suitable for the analysis of small sample data.Dynamic cognitive diagnosis is also known as knowledge tracing(KT)that can diagnose students' current skills in real time according to students' historical answer records.The existing methods are domain-specific(school,grade),so these models do not have the ability to transfer from one domain to another.In order to solve the problems of static and dynamic method respectively,this study proposes two frameworks:Cognitive Diagnosis based on Deep Item Response Theory,Dynamic Cognitive Diagnosis based on Domain Adaptation.Specifically,for the first one,a knowledge point mastery degree vector is initialized for students,and each dimension represents the mastery degree of the corresponding knowledge point.Then,the deep learning model is used to investigate the parameters of learning project reflection theory from the mastery degree vector,topic text and topic.Finally,the item response function of item response theory is used to predict students' scores.For the second one,by minimizing the reconstruction error on the source domain and target domain data sets,a self-encoder is pre-trained together to select the source domain data that is helpful for the target task.Then,according to the maximum mean difference(MMD)distance measure,the distance between knowledge state domain distributions is reduced.Finally,we use fine-tuning technology to solve the problem of inconsistent output of different domain models.We conduct many experiments on some data sets and use the results to show that the model proposed in this work is fine,and achieves good results,which fully verifies the effectiveness of the model.