| Educational value-added assessment(VAA)evaluates the “net output” of the object in a certain period,which restores the diagnostic and incentive function of educational assessment and makes the assessment more objective and fair.Value-added assessment method(or model)is the quantitative means of value-added assessment,Which is the core of value-added assessment.At present,most domestic and foreign researchers conduct value-added analysis based on description statistics and statistical regression models,rarely apply machine learning to value-added assessment.In other fields,machine learning-based analysis methods show the advantages of accurate prediction results,adaptive,and not requiring too many statistical assumptions.With the development of education informatization,the combination of education,artificial intelligence and big data has become an inevitable trend.In view of this,this study applied the machine learning algorithm CART(classification and regression trees)and BP(back-propagation)neural network to value-added assessment,and compared with the standard classification method(ZP)based on description statistics and the multi-layer linear model(HLM)based on statistical regression model to investigate the performance in simulated data and empirical data,specifically:The first research is a simulation research,the simulation data under six experimental conditions were generated based on the random intercept model,and the value-added analysis of ZP,HLM,CART and BP was realized in R,and the authenticity of the value-added results of the four methods under six experimental conditions was investigated.The second study is an empirical study.The data comes from a high school test data in M city,with two value-added analyses: Test 1-Test 2 and Test 2-Test 3.The consistency and difference of the value-added results of the four methods in the subject of language and number are examined.Finally,the results are multi-factor analysis of variance and correlation across time and disciplines,and put forward suggestions on the selection and application of value-added evaluation model.The study conclusions are summarized as follows:(1)In the simulation research,the accuracy ranking of the value-added analysis was HLM> CART> BP> ZP,the correlation with HLM was BP> CART> ZP,and the stability ranking was HLM> CART> ZP> BP.(2)In the empirical research,the consistency ranking with HLM was CART> ZP> BP,and BP showed high and low correlation with HLM in the empirical data with instability.(3)CART and HLM are highly correlated and stable,while BP and HLM are less correlated than the former and lack of stability.(4)Taking test 1-Test 2 Chinese subject as an example,the four methods of student value-added assessment and result assessment are both correlation and difference.The results of the multifactor ANOVA will also depend on the method.The cross-time correlation and cross-subject correlation of students are low,indicating that the value-added of the same student in different times and that of different disciplines are unstable,and the value-added analysis should be carried out according to the specific situation.Overall,the CART method performed well in both simulation and empirical research and is highly correlated with traditional HLM,while BP neural network algorithms which have been shown to perform well in other fields performed poorly in this research.In the future,the data structure,variables and methods of value-added assessment should be further deepened researched. |
