Graph Alignment Algorithm And Applications Based On Deep Learning

The graph alignment task aims to discover the correspondence between nodes in two graphs,and is one of the important tasks in graph data mining.It has a wide range of practical applications,such as social network user matching,knowledge graph fusion,etc.In recent years,with the development of deep learning technology,graph alignment techniques based on node representation have made great progress.These techniques learn node representations from graph data,and then determine the node correspondences by calculating the similarity of the node representations.However,current node representation-based graph alignment algorithms suffer from problems such as excessive focus on local structure,poor scalability,and poor performance in weakly supervised environments.In this paper,we conduct in-depth research on these issues,propose innovative solutions,and verify the effectiveness of the proposed approach through experiments.The main contributions of this paper are as follows:(1)Aiming at the attribute graph alignment problem based on graph neural network,the relation between spectral alignment algorithm and graph neural network is proved theoretically.A deep graph alignment network(DGAN),consisting of deep neural network(DNN)and graph convolutional neural network(GCN)modules,was proposed to increase the difference between node representations and reduce the inconsistency between center node similarity and neighborhood similarity,and to learn node representations with alignment awareness under the guidance of attribute supervision module.Extensive experiments on public benchmark datasets validate the good balance between alignment accuracy and efficiency of DGAN.(2)A trainable cross-graph embedding graph alignment(CEGA)model is proposed to solve the current unsupervised plain graph alignment problem based on node representation.It is based on extending the random walk algorithm from a single graph to two graphs and generating node representations to capture both structural proximity and positional proximity.By treating node proximity as a trainable parameter,CEGA can automatically learn the switching probability parameter of the node to adjust to the current alignment task.Extensive experiments on public benchmark datasets validate that CEGA is superior to the current unsupervised plain graph alignment algorithm.(3)To solve the unsupervised plain graph alignment problem in terms of subspace alignment,an alignment algorithm based on multi-order matching neighborhood consistency(MMNC)is proposed.Only a few pseudo-alignment seeds are used to represent the proximity-aware nodes of two graphs in vector space alignment.By extending the matching neighborhood consistency(MNC)based on Jaccard similarity to vector space,a embedding-based matching neighborhood consistency(EMNC)is proposed.Then by minimizing the loss function based on EMNC,the orthogonal transformation matrix between two sets of node representations is efficiently and accurately approximated using finite pseudo-alignment seeds.Extensive experiments on public benchmark datasets validated MMNC’s superior alignment accuracy and efficiency to the state-of-the-art methods at the time.(4)A position enhanced entity alignment model(PEEA)was proposed to solve the entity alignment problem under weak supervision.In addition to integrating structural information and relational information,the model also introduced position information.Specifically,in order to make full use of the limited alignment seeds,a new position encoding method was introduced to consider alignment seeds and relational frequencies from a global perspective,and then the proposed position encoding was input into PEEA as an additional entity feature to assist alignment search.Extensive experiments on public datasets show that PEEA performs better than other entity alignment models in weakly supervised environments.