Knowledge tracing models take the learner's historical learning behavior data as input, describe the learner's concept mastery state through the representation of the concept, and thus predict the learner's future learning performance. However, in terms of the epitaxial representation of concepts, the conceptual epitaxial information in the current knowledge tracing research is limited to the scope of first-order associations, and cannot represent the epitaxial information of the concept above the first order. To solve this problem, the proposed method first used the graph structure to describe the conceptual connotation information and its interrelationship. Secondly, it used the pooling operation of the graph neural network to extract the epitaxial representation of the concept, which ensured that the epitaxial information of the concept could be derived from the multi-order correlation relationship. Then integrated with the connotation representation of the concept. And finally it predicted the learner's future answers. In order to verify the effectiveness and efficiency of the model, four relevant mainstream knowledge tracing models were selected as comparative models, and experiments were conducted on four commonly used knowledge tracing datasets. The results show that the proposed model has achieved certain advantages in several evaluation indicators, which shows its effectiveness, and in terms of model performance, the proposed model requires the least number of iterations to achieve the optimal evaluation index, which shows its efficiency. In terms of practical application, based on this model, an intelligent learning platform is realized, which judges and predicts the future answers of learners in the teaching process of three offline courses, and achieves better performance than other knowledge tracing models.