| The development of the Internet has brought a serious problem of information overload.As the main means to alleviate this problem,recommendation system has become a research hotspot in academic and industrial fields.Collaborative filtering algorithm is the most widely used among the algorithms in recommendation systems,but there are still some problems,such as poor effect when data is sparse,difficult to explore highorder interactive information and so on.Based on graph neural network,this thesis designs two collaborative filtering recommendation algorithms based on graph convolution neural network to alleviate the above problems and achieve better recommendation effect.Starting from exploring the potential attributes of users,this thesis proposes a recommendation algorithm CD-GCN based on community detection and graph neural network.In the algorithm design,this thesis first explores the distribution law of user and item interaction in the real datasets and establishes a weighted user-user interaction network constructed through user and item interaction information,on which community detection is carried out to find the potential characteristics of users.Then the item attribute information is introduced as auxiliary information,and a collaborative filtering recommendation algorithm based on community detection and graph neural network is designed by using graph convolution neural network,and its effect is verified in the real recommendation system datasets.In the collaborative filtering algorithm,the attributes of the item are used as nodes to build a three-part graph with users and items,which can also be used in the recommendation algorithm based on graph neural network.This method takes attributes as nodes for graph representation and learning,which is suitable for the scene without artificial feature engineering.Based on this method,this thesis proposes a lightweight graph neural network recommendation algorithm light-A2 GCN based on attribute and attention mechanism.Different from the existing methods,this thesis optimizes the GCN module,discards the redundant nonlinear activation part,and constructs a pre model to initialize the attribute node,which improves the performance of the algorithm,and also verifies the effect of the algorithm in the real datasets. |
