Study Of The Multi-view-based Graph Representation Learning For Recommendation

With the development and popularization of information technology and Internet application,modern society has entered the era of information overload.The explosive growth of information and the growth of users make it difficult for information producers to recommend accurate information to users who need it,and it is also difficult for users to select interesting content from the vast amount of information.Recommendation system is the technology dealing with the huge information gap and information demand.It is recommended that the system filter the information,learn the user’s past behavior habits,and provide users with personalized content that meets their preferences.Aiming at the research and development of the existing recommendation system,this paper makes some explorations:The typical recommendation algorithm based on collaborative filtering is based on Euclidean space structure,and it is difficult to fully mine the attribute information and structural relationship information of nodes.The data in the recommendation system is unstructured interactive data,which naturally conforms to the graph data structure.In recent years,graph neural network has been widely used in recommendation.The research core of graph neural network is to aggregate neighborhood information through information transmission mechanism,so as to map the nodes in the graph into low-dimensional vectors with rich information representation.Traditional information transmission networks need to overlay multi-layer aggregation networks to obtain highorder interactive features,which leads to the difficulty of convergence in the training process and its application in large-scale data sets.This paper presents a recommendation model of attention mechanism based on extreme graph convolution.In this model,the limit graph convolution algorithm is used to fully mine the data relationship in the user-project interaction graph,and the high-order interactive embedded representations are extracted from multiple views,and then the information is weighted and aggregated by attention mechanism,which improves the accuracy and diversity of recommendation results.In this paper,compared with the related algorithms on three real data sets,the experiments show that the performance of this algorithm is better than that of the existing models.The data in recommendation system can be regarded as graph structure data in essence,which makes the application of graph neural network(GNNs)in recommendation system more and more in-depth and more extensive.At present,the research results mainly focus on interactive bipartite graph of user-project,which realizes the layer-by-layer transmission of messages and extracts high-order interactive information through multilayer graph neural network.This kind of graph is a simple graph without direction information and symbol information.The information is based on the first-order neighborhood aggregation information,and the central node is directly aggregated with adjacent nodes.The outstanding application result of this research is Graph Attention Model(GAT).However,the GAT model is to extend the attention mechanism to the user-project interaction graph.There is no aggregation mode for interactions with negative relationships,and there is no effective information representation for directional information.With the vigorous development of Internet applications with social networking functions such as Tik Tok and Xiaohongshu,users are increasingly recommending social networking Internet applications,and the data in the system has more and more attributes of social relations and personal likes and dislikes.With this development process,the data structure of recommendation system needs to use signed and directional graph structure to fully express data information.According to sociological theory,this paper selectively models social relations and personal preferences in social networks,and puts forward SDMV model,which extends graph volume(GCN)and attention mechanism(GAT)to graph structures with symbols(users’ likes and dislikes)and directions(social relations).In this model,the symbolic information,the interactive direction information and the steady triangle model between interactive nodes are modeled from the multi-view perspective,and the structural relationship in the data set is fully mined,while the unstable state which is prone to noise interference is eliminated.Experiments on three real data sets show that SDMV performs better than the existing related models.