Research On Deep Knowledge Tracking Model Based On Students Behavior Characteristics

Education data mining refers to the application of theoretical knowledge and technology from multiple fields such as education,computer science,psychology,statistics,etc.to solve practical problems in the field of education.The arrival of the information age means that the impact of each field is unprecedented.Nowadays,the diversification of online education not only brings abundant learning resources to students,but also means an increase in their educational data.In order to provide personalized learning recommendations for students,improve the application value of artificial intelligence technology in the field of education,and liberate teachers from the heavy mechanical working environment,this article models students’ personal historical answering behavior sequence and knowledge point mastery,To construct a more effective knowledge tracking model and provide a better reference model for the analysis of educational data.First,the paper combs the existing knowledge tracking models,analyzes the most comprehensive data set of current education data,and digitizes the data such as stability test,missing values,outlier and other characteristics.Compare and analyze the model processing results by using multiple basic knowledge tracking methods.By comparing the AUC values of Bayesian network student knowledge tracking model(BKT),dynamic key value storage network model(DKVMN),item response theory model(IRT),and deep knowledge tracking model(DKT)under different data sets,we can see that although each method has advantages and disadvantages,the DKT model is better overall,Therefore,DKT was ultimately chosen as the starting point method for this paper’s research.Secondly,in full consideration of the impact of students’ answering behavior sequence on students’ learning process,the paper uses random forest feature extraction to sort out the behavior characteristics in the data set by sorting out the current knowledge tracking model,and obtains the behavior characteristics with the highest importance.Then,the PCA dimension reduction idea is used to find the principal component factors after feature extraction.The model is embedded into the input layer through neural network learning to capture the relationship between different features,and build an external weighted depth knowledge tracking model for random forest feature extraction(DKT-RFPC).Finally,considering the behavioral characteristics of students who rank high in feature importance,excluding objective features,three features are obtained,namely the total number of prompts for students to answer questions,the number of attempts to answer questions,and whether the answers are correct.In the current knowledge tracking model,a single feature strategy,two feature strategies,and multidimensional feature strategies are combined to introduce attention weights in the prediction stage,Output deep knowledge tracking models for attention weights under six different policy combinations(DKT-AHT,DKT-AHC,DKT-AHTHC,DKT-AHTC,DKT-AHCC,DKT-ATCC),to compare and verify the potential relationships between different model policy combinations and obtain different learning structures.