Attribute graph group discovery method based on maximized mutual information and graph neural network

Abstract

The invention provides an attribute graph group discovery method based on maximum mutual information and a graph neural network. The method is characterized in that the method comprises steps of carrying out the representation learning of a to-be-processed matrix through a pre-trained graph neural network, obtaining a preliminary node representation, and carrying out the mutual information calculation of a to-be-processed attribute graph, and obtaining a global mutual information value; dividing the preliminary node representation to the centers of a plurality of groups by using soft clustering to obtain an allocation matrix; carrying out modularity and mutual information calculation in the to-be-processed attribute graph on the original group according to the allocation matrix to obtain amodularity value and group mutual information; and calculating total loss according to the modularity value, the group mutual information and the global mutual information value, and iteratively updating the graph neural network through gradient return according to the total loss till a group discovery result is obtained. According to the method, the end-to-end updating graph neural network doesnot need to be realized step by step, the node attribute relationship can be better captured, and a group discovery result with higher accuracy is obtained.

Description

technical field [0001] The invention belongs to the technical field of big data, and in particular relates to a method for discovering attribute graph groups based on maximizing mutual information and a graph neural network. Background technique [0002] A network graph consists of several nodes and links connecting these nodes. It exists widely in real life and can be used to represent the connection between objects. Representing the structure of network graphs and node attributes in the form of vector data has attracted the attention of many researchers, especially representing the nodes contained in complex and irregular network graphs as vectors of equal length and low dimensionality is the focus of research. As the basis of machine learning, this vector can make machine learning tasks such as node classification, node clustering, anomaly detection and link prediction perform better. [0003] Currently, Community Detection is widely used in node clustering tasks. Among...

AI Technical Summary

Problems solved by technology

However, when this method is reconstructed, only the edge-related information is used as the optimization goal, and it lacks the modeling of the attribute characteristics of the attribute graph itself.

Moreover, the iterative update of the graph neural network generally requires supervision information, and it is very difficult to set appropriate training objectives to deal with unsupervised problems such as group discovery.

Images