Research On Graph Neural Network Recommendation Method Based On Supervised Contrastive Learning

The recommendation system can help users dig out potentially interesting items from massive candidate data,alleviate the problem of information overload,and improve the user experience.Graph neural networks can extract high-order interaction features by aggregating the neighborhood information of the node,so they are widely used in recommendation systems.However,recommendation systems based on graph neural networks still have problems such as data sensitivity,slow model convergence,and insufficient recommendation performance.Based on the above limitations,this thesis conducts an in-depth study on recommendation methods based on graph neural networks.First,starting from contrastive learning,this thesis investigates how to use contrastive learning to improve the performance of graph neural network recommendation models.Furthermore,starting from the loss function,this thesis studies how to improve the existing recommendation loss function to improve the performance and training efficiency of the recommendation model.Specifically,the main work of this thesis is as follows:(1)Aiming at the problem that existing methods do not fully combine target tasks and are sensitive to data,this paper proposes a new Supervised Contrastive Learning Paradigm based on Graph Neural Network(SCL).SCL builds supervised information for representation learning based on the objective function of the recommendation task.The supervised contrastive learning loss guided by supervisory information makes the nodes with similar interaction history close to each other in the representation space,which is beneficial to the improvement of recommendation performance.In addition,a new data enhancement method called node replication is proposed in the SCL paradigm,which enhances the robustness of node representation and the ability to express interest tendencies,making the recommendation results more diverse and further improving the performance of the recommendation system.Indicates ability.Comparative experiments and ablation experiments on standard data sets such as Gowalla,Yelp2018 and Amazon-Book show that the SCL paradigm proposed in this paper is superior to other algorithms in the current recommendation system field in terms of Recall,NDCG and other indicators,and has potential application value.(2)Aiming at the slow convergence speed of existing recommendation algorithms and the need for further improvement in recommendation performance,this paper proposes a new loss function called Supervised Personalized Ranking based on prior knowledge(Supervised Personalized Ranking based on prior knowledge,SPR).The recommendation algorithm based on SPR loss makes full use of the prior knowledge of the original data to construct a supervised signal,and improves the traditional recommendation loss function by using the supervised signal to establish additional user-user relationship pairs,so as to more effectively analyze the user’s interest trend.Modeling improves the ability to recommend the representation of model nodes,thereby further improving the recommendation performance.In addition,by incorporating more prior knowledge,the recommendation algorithm based on SPR loss simultaneously introduces user-item relationship pair and user-user relationship pair information,which greatly accelerates the convergence speed of the model and significantly reduces the practical application of complex models.difficulty.On standard data sets such as Gowalla,Yelp2018,and Amazon-Book,this paper systematically analyzes the recommendation performance of SPR loss combined with various backbone networks.The comparison experiments with existing loss functions and the results of ablation experiments show that the method proposed in this paper is better than Recall,Superiority in indicators such as NDCG and number of iterations.