Research On Deep Knowledge Tracing Method Via Enhancing Question Embedding

Knowledge tracing(KT)is one of the important research direction in the field of educational data mining.It aims to mine the potential learning rules from the information of students’ historical learning trajectory,and then the future performance can be predicted.Through knowledge tracing,online platforms can estimate the knowledge state of learners in real time and make personalized learning resource recommendations.At the same time,further analysis of knowledge state can also build a knowledge map,helping the platform to formulate a more reasonable teaching plan.Existing knowledge tracing methods have mainly considered the interactions between students and questions,but fail to model more intrinsic relations hidden in the interactions between questions and concepts,or only utilize the folded question-concept bipartite graph information.Thus,these approaches cannot capture the relations contained in the student-question-concept interaction at a fine-grained level when modelling question embedding,which will affect the effectiveness of knowledge tracing.To this end,this thesis investigates deep knowledge tracing via enhancing question embeddings from two main perspectives: the calibrated Q-matrix perspective,which mitigates the subjective bias in Q-matrix,and the pre-training question embedding perspective,which is rich in potential semantic information.The following results was achieved:(1)We devise a novel Calibrated Q-matrix-based Knowledge Tracing(CQKT)framework to track knowledge proficiency of students dynamically in KT.To be specific,for the original Q-matrix,we primarily strive to capture the high-order connectivity between questions and concepts to obtain potential concepts of each question by utilizing graph convolution network.Then,three pairwise Bayesian methods equipped with potential concepts are adopted to refine and calibrate the raw Q-matrix so that the subjective tendency of the Q-matrix defined by domain experts can be weakened.After that,the embedding of each question aggregates the calibrated Q-matrix with a KT model to trace students’ knowledge states.(2)We present a novel Semantic-Enhanced Question Embeddings Pre-training(SEEP)method,concentrating on decomposing the underlying semantic information in the interactions and further fusing the information of questions and concepts under different decomposed perspective semantics to obtain semantic-enhanced question embeddings for improving the performance of KT methods.Concretely,from two semantic perspectives(co-occurrence and answer-agreement),we first mine relations between questions(and concepts)as the prior.Thereafter,by extracting question-concept bipartite graph from the student-question-concept interaction,we utilize a two-level attention aggregation mechanism with the prior information to attain node embeddings of questions and concepts.Furthermore,for obtaining the final question embeddings,we solve a joint optimization problem by taking question difficulty constraint and questionconcept relation structure constraint into account.Finally,extensive experiments on three real-world educational datasets demonstrate that the two methods outperform the baseline methods compared in the thesis,validating the effectiveness and reasonableness of the models.