Research On Multi-Relationship Recommendation Model Based On Graph Convolution Network

In the context of digitalization,related research in the field of recommendation systems has received widespread attention in all walks of life.Because Graph Convolution Network(GCN)can use the cooperation signals of high-order neighbors to learn the embedding of users and items better,it has been widely used in recommendation systems.In the current GCN-based multi-relational recommendation model,the embedding learning of user nodes will be interfered by high-order neighboring users whose interests are not similar to them,resulting in users with different interests getting similar embeddings after multi-layer graph convolution,resulting in the over-smoothing problems and reducing the accuracy of recommendations.In addition,the existing multi-relationship models often directly follow the functions such as feature transformation and nonlinear activation used in standard GCN,and do not deeply study whether the above operations will have an impact on the collaborative filtering task,and the differences between different behaviors are often ignored in the process of fusing multi-relationship embeddings,so that the multi-behavior information cannot be fully utilized for embedding learning.In this paper,the following work has been carried out around the above problems in the multi-relationship recommendation algorithm:1.Aiming at the over-smoothing problem in the high-order propagation of the existing multi-relational recommendation model,this paper uses the subgraph generation module to assign users with similar interests and their connection items to the same subgraph,and then when the subgraph carries out high-order propagation,users can make full use of the information of nodes related to their interests to carry out embedded learning,which effectively alleviates the over-smoothing problem in the existing model and improves the accuracy of recommendation.2.Aiming at the problem that existing multi-relationship recommendation models have multiple redundant operations in graph convolution,this paper reduces the number of parameters in the model by removing the feature transformation and nonlinear activation function used in conventional GCN,and the improved lightweight graph convolutional network can improve the computational efficiency of the model and improve the recommendation effect.3.Since the GCN-based multi-relationship recommendation model ignores the different influences of different behavior types on user preference learning in the multi-relationship fusion stage,this paper adds correlation behavior coefficients to better use multi-behavior information learning embedding representations in this process,and proves the effectiveness of the design by comparing the two dimensions of unified weights and different weights.Aiming at the above problems,this paper proposes a multi-relational recommendation model based on user interest-aware(IMRRM).The effectiveness and robustness of the model are verified by experiments on two large-scale real datasets,Beibei and Taobao.The experimental results show that the HR10 and NDCG10 indexes of the IMRRM model on the Beibei dataset are improved by 1.98% and 1.58% compared with the traditional method,and by 1.49% and 1.81% on the Taobao dataset,respectively,which has good performance.