Graph neural network recommendation method and system fusing social relation and semantic relation

Abstract

The invention belongs to the technical field of graph neural network recommendation, and discloses a graph neural network recommendation method and system fusing a social relation and a semantic relation, which utilize a simple graph neural network recommendation model to recommend a graph neural network based on user social information, score interaction information of users and items and label information of the items. Constructing the information of the non-graph structure into graph information; clustering the labels of the items by using the label information of the items; and calculating the relevancy between the items and the label clusters, carrying out cosine similarity calculation by utilizing an item-label cluster relevancy matrix, and when the similarity is greater than a certain threshold value, adding a trust relationship between the two items to form a semantic relationship graph structure of the items. According to the method, the user-project score bipartite graph is used as a basis, other isomorphic graphs are used as supplements, so that the purpose of using more neighbor nodes for aggregation is achieved, the problem of data sparsity is solved, and tests prove that the proposed model has higher recommendation accuracy than GrpahRec.

Description

technical field [0001] The invention belongs to the technical field of graph neural network recommendation, in particular to a graph neural network recommendation method and system that integrates social relationships and semantic relationships. Background technique [0002] At present, the most traditional recommendation algorithm is a collaborative filtering algorithm that recommends by calculating the neighbors of users or items. Because the original algorithm has limitations, many researchers have been attracted to improve it. In order to solve the problem of data sparsity and cold start, with the research of context information, the traditional recommendation algorithm using context information is to use context information as a supplement when calculating similarity. For example, the existing literature uses label information as a supplement for collaborative The similarity calculation in filtering can solve the problem of inability to make recommendations due to lack ...

AI Technical Summary

Problems solved by technology

[0006] (1) Traditional recommendation algorithms using contextual information can only incorporate one type of contextual information, and cannot obtain comprehensive recommendation results; items that users have not rated should not be used as user feature values, so the forcedly constructed user -The item scoring matrix is ​​prone to the problem of data sparsity, which leads to the problem of low accuracy of the recommendation algorithm because of the inability to calculate the similarity due to data sparsity

[0007] (2) In the existing score prediction method based on graph neural network, only one kind of graph information is used, and the own node information is ignored when the neighbor node information is aggregated, and it is assumed that the influence of all neighbor nodes on the user is equally important No context information is added to the item model, and in the learning process of the user latent factor vector, the difference in the degree of influence of the two graph information on user preferences is not considered

[0008] (3) In the existing method of using the dual graph attention network, the context information of the item is also established based on the social information of the user, rather than the context information of the item itself; the existing model only considers one kind of context information, even if There are many considerations, such as extending the user's social influence to the project domain, and it is only based on a kind of contextual information.

The emergence of the graph neural network can easily integrate various graph information, but most of the context information is not graph information, so it is very difficult to integrate different context information into the graph neural network.

Images