Research Of Knowledge Tracing Model Based On Dynamic Key-Value Memory Network

In recent years,driven by artificial intelligence technology,the computer-aided teaching system,which is different from traditional education,has developed rapidly.Knowledge tracing is a mature and challenging problem in this field.Its main task is to use the interaction data of students and the contents of the learning system to model the knowledge state of students,and then predict how students will behave in the future interaction.Although there are many models for knowledge tracing tasks,the current research still has the following problems:First,from the perspective of solving mathematical problems,the modeling of students’real learning process is not sufficient,and the influence of various factors such as forgetting is ignored,which limits the effect of knowledge tracing.Second,the complexity of student learning data is not fully considered when modeling the learning process,while the information of student learning data is not fully utilized.Third,the similarity’ of questions is usually ignored,which means long-term dependencies in student-answer sequences cannot be effectively captured.In view of the problems existing in the above knowledge tracing model,the main work and innovations done in this thesis are as follows:First,this thesis proposes the KTIFM(Knowledge Tracing with Improved Forgetting Mechanism)model with improved forgetting mechanism.Firstly,in terms of time factor,the sequence time interval and repetition time interval related to students answering questions are defined,and these two features are introduced into the knowledge tracing model through feature fusion,so that the model can consider the time factor when calculating the amount of students’ knowledge forgetting.Secondly,in terms of knowledge state,this thesis improves the internal structure design of DKVMN(Dynamic Key Value Memory Networks),so that the model considers the influence of the current knowledge state when calculating the amount of students’ knowledge forgetting and knowledge growth.The forgetting factor is also introduced into the prediction process of the model.Finally,in order to verify the effectiveness of the improvement,experiments were carried out on the data set ASSISTments2009 and the data set SmartEdu2022.Compared with the standard DKVMN model,the experimental results show that the AUC values of the improved KTIFM model are improved 2.2%and 1.1%respectively,indicating that the knowledge tracing effect of the improved model has been effectively improved.Second,this thesis proposes the Sub-KTIFM model(KTIFM model based on Sub-LSTM)based on Sub-LSTM(improved LSTM based on Subsequence).Firstly,the similarity of the questions is measured based on the triangular membership function,so that the answer sequence can be divided into different subsequences.Secondly,the Sub-LSTM mechanism is designed to realize the selective connection of the input sequence.Finally,the Sub-KTIFM model is composed of the Sub-LSTM mechanism and the KTIFM model.The use of the Sub-LSTM mechanism enables the model to effectively construct the long-term dependencies of students’ answer sequences.The experimental results on the two datasets show that the AUC values of the Sub-KTIFM model are improved by 3.9%and 1.6%,respectively,which verifies the effectiveness of the Sub-KTIFM model.