Research On Knowledge Tracing Based On Dual Graph Model Ensemble

Education plays a pivotal role in driving social progress and development.It fosters critical thinking and innovation,enhances individual and societal competitiveness,and contributes to the holistic development of humanity,while also promoting equity and justice.In recent years,thanks to the rise of the Internet and big data technology,largescale online learning platforms such as MOOCs have developed rapidly,and the number of users on the platform has increased sharply.How to take advantage of the number of users of the online platform and the Internet big data technology to improve teaching quality and provide students with adaptive and personalized guidance has always attracted educators and educational data mining scientists.Knowledge tracing aims to realize real-time evaluation and tracing of students’ knowledge level based on students’ historical learning records,to accurately judge students’ mastery of different knowledge concepts,bring possibilities for personalized education,and play a vital role in intelligent education.With the development of Deep Neural Networks,especially Recurrent Neural Network,because of its strong feature extraction ability,it is suitable for modeling complex learning processes.Compared with traditional methods,the knowledge tracing method based on deep learning has a great improvement in performance.The existing methods introduce relationship information to improve knowledge tracing.Although there is some progress,they usually only use a single type of relationship.At the same time,the potential benefits of integrating different models for knowledge tracing have not been fully developed.Therefore,this paper studies the knowledge tracing of double-graph model fusion,and the work is as follows:(1)There are many relationships in knowledge tracing,such as between exercises,knowledge concepts,exercises and corresponding knowledge concepts,and students’ learning interactions.The current knowledge tracing model only considers the single relationship information between exercises or knowledge concepts,and does not fully mine the relationship information between students’ learning interactions.Therefore,this paper proposes a knowledge tracing method based on a dual graph neural network,including concept association hypergraph and directed transfer graph,which are respectively used for the higher-order relationship between modeling exercises and knowledge concepts,as well as between student interactions.Through hypergraph convolution neural network and directed graph convolution neural network,the interactive feature representation containing rich relationship information is learned,thus improving the prediction performance of knowledge tracing model.The experimental results on three benchmark data sets show that the method is effective and powerful in feature extraction.(2)A single model is often difficult to model complex learning and cognitive processes,so the performance of knowledge tracing can be further improved through model fusion.At present,model fusion in knowledge tracing research is usually achieved by simple weighted summation or splicing.At the same time,the existing methods only optimize the model according to the students’ performance in a single exercise at each moment.To solve this problem,this paper proposes a knowledge tracing method based on knowledge distillation model fusion.This method mainly includes two modules:interactive sequence modeling and online knowledge distillation.First,different sequence models are used to model students’ learning interaction sequence,and gating mechanism is introduced to measure the importance of different models,and online knowledge distillation is used to form a stronger fusion model.Use the fusion model to predict students’ performance in all exercises as an additional monitoring signal to guide the optimization of a single knowledge tracing model and improve the representation ability of the model.The experimental results show that the model fusion method based on knowledge distillation proposed in this paper can achieve better performance in knowledge tracing tasks than the simple weighted summation or splicing method.