Graph embedding method based on adaptive graph learning

Abstract

The invention discloses a graph embedding method based on adaptive graph learning. The graph embedding method comprises four steps of constructing a graph auto-encoder framework, performing Laplace embedding, performing adaptive learning on an adjacent matrix and performing iterative updating solution. A two-layer graph convolutional neural network is adopted in the coding layer part of the graph auto-encoder frame, and the reconstruction loss of an adjacent matrix is formed in the decoding layer part; the Laplace embedding part is used for embedding a Laplacian matrix into a potential space, so that a sample point can be more accurately mapped to a projection subspace; the adaptive learning of the adjacent matrix is divided into three steps: 1, the fixed number of node neighbors is not adopted any more, but normal distribution is obeyed, so that variables which can be obtained through the adaptive learning are formed; 2, a threshold value for stopping iteration is set, and updating is stopped when the number of iterations is greater than the threshold value; 3, part updating of the adjacent matrix is expressed by the formula in the specification; and finally,a model solving method is provided in an iterative updating solving part. The method is high in robustness and wide in application range, and the application range of the graph auto-encoder is greatly expanded.

Description

technical field [0001] The invention belongs to the technical field of data mining, and in particular relates to a graph embedding method. Background technique [0002] Graph structures are found naturally in various real-world applications, such as social networks, word co-occurrence networks, and communication networks. Research on graph neural networks is closely related to graph embedding or network embedding. The purpose of graph embedding is to represent network nodes as low-dimensional vectors while preserving network topology and node content information. The purpose of graph neural networks is to solve graph-related tasks in an end-to-end manner, which extracts high-level representations. The relationship between graph neural network and graph embedding can be understood as: graph neural network can be designed for a variety of tasks, in which the graph autoencoder framework can be used to solve the graph embedding problem, and graph embedding includes various oth...

AI Technical Summary

Problems solved by technology

[0004] But both methods have their limitations

First: These methods still use k-nearest neighbors (kNN) to initialize the adjacency matrix, which has the disadvantage of fixing the k value; second: both methods are aimed at supervised graph neural network frameworks rather than unsupervised graph autoencoders frame

Images