Research On Knowledge Tracing Based On Knowledge State Evolution Representation

Knowledge Tracing(KT)aims to predict students’ knowledge state and future performance based on the past exercise.KT is an important student modeling method and the key technology of intelligent education.According to the representation of knowledge state,the existing KT models are divided into the probabilistic KT and deep learning KT models.The two have achieved well prediction performance.However,the probabilistic KT models use the same state parameters model all students’ knowledge state,which ignores the individualization;the deep learning KT cannot provide directly interpretations by the hidden vectors.To solve the problems,the paper proposes a Student Vector Representation Networks,and defines the state vector to represent knowledge state.The state vector combines the advantages of state parameters and hidden vectors.and it can model the individualization and has directly interpretations.Furthermore.the model does not restore the knowledge state variation,so the paper further proposes a Student Vector Representation Evolution Networks,which defines the vector evolution term to model the slowly knowledge state variation.But the two existing KT models and the two proposed models still exist problems.Firstly,the probabilistic KT models is specific for a skill;the state vector only represents the knowledge state of the answered skill.So they cannot model skill correlation.Secondly,the deep learning KT models summarize the knowledge state of all skills,so they hardly model the knowledge state of a skill.To solve the problems,the paper further proposes a Student Matrix Representation Networks,and defines a state matrix to represent knowledge state.The row in the state matrix is the knowledge state of specific skill,and the row correlation is the skill correlation.In the same way,the paper also proposes a Student Matrix Representation Evolution Networks,and defines the matrix evolution term to model the slowly knowledge state variation.In summary,this paper proposes four KT models.The best is the Student Matrix Representation Evolution Networks,which not only accurately represents the students’knowledge state,but models the slowly knowledge state variation.In experiments,our model achieves improvements in prediction performance and interpretations.