Research On Intelligent Tracing Method For Student Knowledge Level

In recent years,the rise of online education has greatly promoted the development of education in the direction of intelligence,and the learning process of learners gradually tends to a more personalized adaptive learning mode.The massive amount of learning data accumulated by online education platforms make it possible for researchers to use data-driven methods to analyze students’ behavior patterns.Analyzing and tracking students’ knowledge states is a key part of the adaptive learning system,which lays the foundation for providing personalized learning resources and planning appropriate learning paths in the future,making students’ learning more efficient.At present,the technology of knowledge state tracking is mainly divided into two modes:static cognitive diagnosis and dynamic knowledge tracing.A lot of research work has been done in this field,but there are still some problems that need to be solved,such as the lack of attention to the impact of collaborative information among students in cognitive diagnosis tasks,the inability of existing knowledge tracing methods based on a single representation of students’knowledge levels to take into account the need for fine-grained state analysis and the role of correlation between exercises simultaneously,and the lack of valid information in the representation of exercises and knowledge concepts.To address these issues,the main work of this paper includes:1.Subgraph Pattern Mining-Based Cognitive Diagnosis To incorporate inter-student impacts into the cognitive diagnosis process,the subgraph pattern mining-based cognitive diagnosis model proposed in this paper designs a subgraph extraction algorithm,which extracts the most relevant local subgraphs from the original studentexercise interaction bipartite graph around the target student-exercise pair,and filters out the rest of the irrelevant parts.The graph neural network is used to perform representation learning to mine the subgraph patterns,so as to introduce the collaborative information among students into the model.Furthermore,the model uses latent representations learned by the autoencoder as auxiliary features for subgraph nodes to compensate for the lack of node specificity due to subgraph slicing.Finally,a cognitive diagnosis module is designed to obtain students’ knowledge status and to make predictions for their future performance.2.Dual-State Knowledge Tracing In order to make up for the shortcomings of the knowledge state representation method in the previous knowledge tracing methods,this paper proposes a new knowledge tracing model,which uses the dynamic routing mechanism to extract the general knowledge commonalities from the original knowledge concepts,and represents students’ knowledge states at the specific knowledge concept level and abstract knowledge commonality level,then predicts students’ future performance by fusing these two knowledge state representations,and update them simultaneously using gated recurrent unit and dynamic key-value memory networks,realizing a multi-level modeling of students’knowledge states.3.Representation Enhancement Methods for Exercises and Knowledge Concepts In order to investigate the influence of the relationship between the exercises and knowledge concepts on the knowledge level tracing method,representation enhancement methods based on mutual information maximization and graph neural network are designed in this paper to incorporate the correlation information between exercises and knowledge concepts into their respective representations.The experimental results show that the effect of the proposed knowledge tracing model on the student performance prediction task is enhanced by attaching the two representation enhancement methods.