Recommender systems are created to help users screen and filter the huge amount of data generated in the Internet, so that users and data can be matched perfectly. This paper establishes the CGSNet session recommendation model from the perspective of session recommendation system on the basis of explicit language model recommendation. The CGSNet model is based on graph convolutional neural network, which extracts the neighborhood information of the user’s session through the fusion of the attention mechanism, and combines the global session representation learning to construct the global graph of the item. The self-supervised comparison learning module is introduced to realize the effective exchange of neighborhood information and global information, and the joint learning objective is designed to optimize the session model recommendation performance. The results show that the P@15 and MRR@15 metrics of the CGSNet model on the Tmall dataset are improved by 15.89% and 19.11%, respectively, compared with the suboptimal model. The effectiveness of the model is verified through multiple types of simulation experiments, which also illustrates the feasibility of comparative learning for optimizing recommendation data in session recommendation.
