Abstract: Aimed at the problem that the difficulty of exercises is not considered in the knowledge tracking model with consistent learning process, a method of quantifying the difficulty based on the results of weighted average answer time and error rate was proposed. Aimed at the problem that the knowledge concepts in the knowledge tracking model with consistent learning process were independent from each other, a network structure was proposed to represent the knowledge concepts. The first-order similarity and second-order similarity between the knowledge concept nodes were calculated based on the joint probability and empirical distribution. The LINE algorithm in the network embedding could retain the advantages of the network structure information. The knowledge concept represented by the network structure was embedded into the low dimensional feature vector. The effectiveness of the method is verified by establishing a model on real and public data.