Heterogeneous Graph Neural Network Based Knowledge Tracing Approach

Rapid development of online learning platform has brought more learning resources and better personalized learning experience to modern education system.But learners’growing needs of personalized learning has made current platform insufficient.Especially in knowledge tracing domain,the problems of how to mine learners’ history interaction data,how to analyze learners’state and how to predict learners’ future performance are all requiring growing attention.Here we aim at knowledge tracing task,integrating knowledge concept and question graph to knowledge tracing models with graph neural network.And we explored a new forgetting behavior based graph updating method.We build a graph knowledge tracing model whose updating process happens in both temporal dimension and spatial dimension.Then we apply heterogeneous neural network to represent questions which is used in following knowledge tracing models.All our work was based on literature survey and theoretical exploration,then we used experimental verification and single variable method to verify the proposed methods.Our work is supposed to improve the existing knowledge tracing models,make full use of the advantages of the diversified data collected by the learning platform,and help the learning platform provide learners with better personalized learning.Many scholars in the domain of knowledge tracing have proposed excellent models and methods,but existing knowledge tracing models still have some shortcomings:Firstly,most existing graph knowledge tracing models focus on the mechanism of graph updating process.They rarely take education domain knowledge,such as learners’ forgetting behavior,into account.Secondly,most deep knowledge tracing models adopt linear model structure,which only focus on the influence of previous interaction on future interaction.They seldom consider the graph structure of knowledge concepts and their mutual influence on the basis of network structure.Thirdly,most existing knowledge tracing models use the question id to identify the interaction targets,but pay little attention to the content of the question itself and the network constructed by question and knowledge concepts.Graph neural network could naturally embed knowledge concept and questions as graph vertices,capturing the network structure better than linear ones.Heterogeneous graph neural network could even embed node contents and graph structure at the same time,which allows us to represent questions with its text content,knowledge concept and difficulty.Here we propose Heterogeneous Graph Neural Network based Knowledge Tracing Approach,HGKTA to complete the question representation and knowledge tracing task.Our contributions are as follows:(1)In order to solve the problem that the graph knowledge tracing model lacks education domain knowledge,we proposed a forgetting behavior integrated graph knowledge tracing updating method.This method mines multi-dimensional forgetting attributes from learners’interaction data,including interaction time interval and repeated practice times,and then integrates the multi-dimensional forgetting attributes into the updating step of the model,so as to simulate forgetting behavior and memorizing behavior as time goes by.Model with this method can update learners state according to degree of forgetting.In the experiment part,benchmark deep knowledge tracing models are used as the baseline to compare and analyze the efficiency of our method.(2)In order to lift the restriction of existing graph knowledge tracing models not being able to mine local features and complex high-level features,we proposed a new graph updating method in temporal spatial dimension.By decomposing graph updating process into two steps:temporal one and spatial one,and then expanding the information transaction range of vertices in graph in spatial dimension,our updating method enables the graph neural network to better obtain the local features and complex high-level features when updating node state.In the follow-up experiments,we compared model with ones without the update method.(3)To solve the problem that question representation not containing text content and graph structure characteristics,we proposed a method embedding the text,knowledge concepts and difficulty of questions with heterogeneous graph neural network.We started by normalizing heterogenous question data like text,knowledge concepts,concept relation and difficulty.Then,the node content and graph structure are integrated into the node representation using the mechanism of heterogeneous graph neural network.Finally,the node embedding is settled and can be applied to downstream knowledge tracing models.In the follow-up experiments,the model is compared with the knowledge tracing models without question embedding.We verified effectiveness of our approaches on Eanalyst dataset(http://study.hub.nercel.com),learning data which is collected in real education scenes and benchmark dataset Assistment.The experiment result shows that:(1)The forgetting behavior integrated graph knowledge tracing updating method can improve model performance to a certain extent.(2)The temporal spatial updating method can improve the performance of the model,and the combination of the two methods can further improve the performance on some dataset(3)The heterogeneous graph question embedding method can greatly improve the result of the model.According to visual analysis of the experimental results,it is proved that the question embedding method using heterogeneous graph is more accurate than other embedding methods.