Research On Knowledge Tracing Based On Graph Neural Networks

As the development of big data and Internet,the online education systems are drawing attention for being the platform of sharing educational resources and the tool of prompting fairness in education.Among various high-level services that online education systems could provide,intelligent education which provides customized tutoring takes an important position.Knowledge tracing is the cornerstone of intelligent tutoring system,therefore improving the performance and efficiency of knowledge tracing task is of great importance to enhancing the intelligent education system.Previous studies on knowledge tracing have made considerable successes,while some issues are still needed to be stressed:1.Previous knowledge tracing models based on graph neural networks failed to capture the relations between nodes representing knowledge concepts and failed to make full use of neighborhood information.2.Despite complicated calculation and high consumption of resources of graph neural network methods,works about slimming down the models for knowledge tracing are short.Considering deficiencies above,this thesis studied the relations between nodes in graph as well as the algorithm for neural network pruning to tackle the aforementioned problems.Specifically speaking:1.A graph-neural-network-based knowledge tracing model employing reciprocal neighborhood relation and incorporating information of second-tier neighbor nodes with a new data processing mechanism and message passing channel is proposed to improve the model performance.2.The graph-based knowledge tracing model is pruned under lottery ticket hypothesis providing masks for both adjacent matrix and neural network,sparsifing the model without impairing its performance,and experiment result validates the existence and effectiveness of pruned network.The experiments of this thesis are based on two real-life education datasets and a synthetic education dataset.Contrast experiments have been made for previous knowledge tracing models that based on graph neural networks and the model proposed in this thesis where the performance results showed the superiority of the latter method.And pruning method based on the lottery ticket hypothesis is conducted to show that it could identify a smaller sub-network with comparable performance of original network,which proves the effectiveness of the method used in this thesis.