Research On Knowledge Tracing Model Based On Data Augmentation

With the popularity of online learning in the educational environment,more and more students are participating in it.It is hard to assess learners’ knowledge mastery in detail with traditional learning approaches.Compared to offline learning methods,online learning has more advantages.On the one hand,online learning platforms retain students’ learning trajectories and provide conditions for tracing different students’ behaviors.On the other hand,online learning platforms can identify learners’ knowledge weaknesses and provide learners with better learning methods.The task of Knowledge Tracing(KT)is to help students learn more effectively by predicting their next level of knowledge based on their historical practice sequences.Nowadays,many achievements have been made in this field,such as Bayesian knowledge tracing and deep knowledge tracing methods.The emergence of graph neural networks in recent years has also spiced up KT.In this thesis,by studying the students’ answer records,the existing graph-based model is enhanced for knowledge tracing.Specifically,this thesis investigates enhancing learners’ answer records to accomplish the knowledge tracing task.In spite of the advances that have been made in the field of KT,existing techniques still have the following limitations:(1)Previous studies have addressed KT only by exploring the observed sparse data distributions,while counterfactual data distributions have been ignored.(2)The current work designed for KT either considers only entity relationships between questions and concepts or between two concepts,while studying multi-entity relationships among students,questions,and concepts simultaneously is missing,which leads to inaccurate student modeling.To address the above issues,this thesis considers the multiple entity relationships of student learning information based on graph neural networks and models counterfactual data from a counterfactual perspective.The research of this thesis consists of three main parts:(1)To address the above limitations,this thesis proposes a counterfactual graph learning method for knowledge tracing.Specifically,to consider multiple relationships among different entities,we first define students,questions,and concepts in a unified graph,and then use graph convolutional networks for representation learning.To model the counterfactual world,we perform the counterfactual transformation of students’ learning graphs by changing the corresponding processing methods,and then use heterogeneous graph neural networks for information transfer and aggregation to model the learners.In addition,this thesis introduces a contrastive learning framework to obtain counterfactual positive samples at the question level using counterfactual transformations to further learn accurate student representations.(2)Considering data enrichment from different levels,this thesis further proposes a student-level enhanced counterfactual graph learning knowledge tracing model.In this model,the construction of student-question-concept graphs in the knowledge tracing model for counterfactual graph learning is first borrowed,and multiple entities of different types as well as multiple types of edges are defined in a single graph.Then,a heterogeneous graph neural network is used to learn the concept,question,and student representations.In modeling the counterfactual world,the model also considers student-level information to enable counterfactual transformation and learner modeling from the student’s perspective.Different from the above model,the model operates directly on the learner during the counterfactual transformation phase based on the student’s learning ability.Moreover,to further learn the accurate student representation,the model is designed for contrastive learning,where observed and counterfactual positive samples(negative samples)are compared to learn the student characteristics.(3)In this thesis,extensive experiments are conducted on three real-world datasets ASSIST2009,ASSIST2012,and Algrbra2006 for the two knowledge tracing methods proposed above.The experimental results demonstrate the superiority of the data augmented knowledge tracing methods proposed in this thesis over the state-of-the-art methods in terms of AUC and ACC,and prove the validity of the counterfactual transformation and contrastive learning.