Unsupervised Adversarial Graph Alignment Based On Graph Embedding And Graph Convolution

Graph alignment,also known as node alignment,aims to find corresponding users in multiple social network platforms,and these users belong to the same natural person in the real world.It is a basic task in social network analysis.In recent years,it has been involved in the field of cross-knowledge graphs and the field of natural language processing.Although traditional graph alignment methods have made some progress with increasing computing resources,most of them rely on training data and test data that require some corresponding nodes,so-called anchor links.The acquisition of these data will involve security issues such as user privacy,etc.If directly labeling the corresponding nodes on various graph network data directly will bring limitations to practical applications.Therefore,how to construct the corresponding nodes is a challenge for graph alignment in practical applications.Based on the characteristics of the graph data itself,this article conducts research on aligning two graphs without any anchor links.Achieving the purpose of not only retaining the characteristic information of the graph itself,but also without anchor links between the graphs.Based on the property of the graph itself,this paper proposes two frameworks to independently solve the problem of unsupervised graph alignment:one is the unsupervised adversarial graph alignment framework based on graph embedding.In a completely unsupervised manner,means there are no ready-made anchor links and no nodes attribute information or other data are available,to conduct cross-domain alignment of the node vector space of the two graphs;the second is an unsupervised adversarial graph alignment framework based on graph convolution,which directly combines node feature vectors and adversarial networks through graph convolution networks to achieve end-to-end training.This paper further expands the two frameworks into incremental ones,which can be retrained through pseudo-anchor links,that is,the output of the model,to obtain unobserved user links.At the same time,this can form positive feedback to further improve the quality and alignment accuracy of the node vector.This paper introduces adversarial training to efficiently label two graphs without any relevant information,effectively reducing the labeling pressure on large-scale data.The incremental framework is to select the high-similarity nodes with the highest confidence to form a node pair.At the same time,the two-way nodes are aligned to select a reliable high-confidence node pair to join the training set,which provides new features to the model and also eases the problem of too little information of the original training data.The scheme proposed in the paper was tested on the public social network databases Last.fm,Flickr,Myspace,and the document databases Cora,Citeseer respectively.Compared with the traditional unsupervised scheme,the alignment accuracy on the graph alignment task has been improved by an average of 40%,while omitting a lot of manual labeling time.In addition to being applied to social networks,the model proposed in this paper can also be used for other graph data alignment applications,such as drug development and image matching.