Unsupervised graph topology transformation covariant representation learning method and device

Abstract

The invention discloses an unsupervised graph topology transformation covariant representation learning method and a device, and relates to the field of unsupervised learning. The method is a universal framework which can be applied to learning graph node feature representation in the GCNN, and graph topology transformation covariant representation is formalized by maximizing graph topology transformation and mutual information between node representation of graphs before and after transformation. At the same time, the present invention demonstrates that maximization of this mutual information can be approximated to minimize cross entropy between graph topology transforms and graph topology transforms estimated from node representations of graphs before and after transformation. Specifically, part of node pairs are sampled from an original graph, connectivity of edges between the node pairs is turned over to achieve graph topology transformation, then graph topology transformation is reconstructed from feature representation of the original graph and the transformed graph, self-training is conducted on a representation encoder, and feature representation of nodes is learned. The method is applied to node classification and graph classification tasks and is superior to the latest unsupervised method.

Description

technical field [0001] The invention relates to the field of unsupervised learning, in particular to a method and device for unsupervised graph topology transformation covariant representation learning. Background technique [0002] Graph is a natural and effective representation of irregular data / non-Euclidean data (such as 3D point cloud, social network, citation network, brain network, etc.). Due to the powerful expressiveness of graphs, more and more attention has been paid to machine learning of graph data, such as the Graph Convolutional Neural Network (GCNN) proposed in recent years. However, most existing GCNN models are trained in a supervised or semi-supervised manner, which requires a large number of labeled samples to learn effective feature representations. Due to the high labeling cost (especially on large-scale graphs), existing methods are difficult to apply widely. Therefore, we need to learn graph feature representations in an unsupervised manner in order...

AI Technical Summary

Problems solved by technology

However, GraphTER only explores the transformation of node features, while the graph topology has not been fully explored, but it is crucial in unsupervised graph representation learning

Images