Research On Graph Alignment Based On Knowledge Transfer

With the development of the mobile Internet,the amount of data has exploded.As a typical data structure,graph can effectively model complex relationships between data,and plays an important role in many fields such as social network,academic network and knowledge graph.However,the information in a single graph is often missing and incomplete.Different graphs need to be interactively fused to fully tap the potential value of the data and maximize the utility of the data.The graph alignment problem,which aims to realize the association of the same entities in two graphs,as the basis of information fusion between graphs,has received extensive attention from academia and industry.In this thesis,two graphs in the graph alignment task are defined as a "domain",and the graph alignment problem can be divided into three key problems according to the number of"domains":single-domain graph alignment problem,double-domain graph alignment problem and multi-domain graph alignment problem.Existing single-domain graph alignment methods have certain limitations when dealing with the scarcity of labeled data,and there are few methods to study double-domain and multi-domain graph alignment problems.In order to solve the above problems,this thesis first studies the single-domain graph alignment problem from the perspective of knowledge transfer,which uses unlabeled data to improve the feature extraction ability of the model and further improves the alignment accuracy.Then,the problem of doubledomain graph alignment is studied,and the external alignment knowledge in the source domain is transferred to the inside of the target domain,thereby improving the alignment accuracy of the target domain.Finally,the problem of multi-domain graph alignment is studied,and domaininvariant alignment patterns are learned using multi-domain data,resulting in robust and generalizable graph alignment models.The innovative research results obtained in this thesis are summarized as follows:(1)Aiming at the single-domain graph alignment problem,this thesis proposes a contrastive learning based single-domain multi-granularity graph alignment method.Existing studies have carried out preliminary explorations on the graph alignment problem,but there are certain limitations when the graph data scale is large and label data is scarce.In order to solve the above issues,this thesis assists alignment by mining the potential hierarchical information of nodes in large-scale graphs,and leverages the idea of contrastive learning to extract useful information from unlabeled data.However,this will face the following challenges:how to extract potential node information in large-scale graphs and how to use unlabeled data to improve model performance.Specifically,this method first proposes a novel multi-granularity graph alignment framework,which divides large-scale graphs into multiple granularities to describe the hierarchical structure of nodes,and utilizes the alignment of coarse-grained graphs to facilitate the alignment of fine-grained graphs.Then,novel intragraph and inter-graph contrastive learning methods are proposed,and the unlabeled data is used to promote the model to learn high-quality node features and alignment patterns to achieve alignment.Finally,this thesis provides a solid theoretical analysis to show the effectiveness of intragraph and inter-graph contrastive learning methods.Experimental results show that the proposed method improves Hits@k by 15.93%and MRR@k by 14.82%.(2)Aiming at the double-domain graph alignment problem,this thesis proposes a transfer learning based double-domain graph alignment method.Existing methods cannot utilize the knowledge in other graph alignment tasks when solving the graph alignment problem.However,solving this issue will face the following challenges:how to design a transferable graph alignment model targeting two graphs in the same domain,and how to transfer the graph alignment model from source domain to target domain.In order to solve above challenges,this method first proposes a novel transferable intra-domain graph alignment model Ego-Transformer,which improves the traditional Transformer,introduces the ego-network attention layer,and leverages the stability of ego network to learn stable alignment features.Then,a novel inter-domain difference elimination model WWGAN is proposed,which improves the traditional WGAN model,introduces a weight mechanism,and uses supervision information to eliminate inter-graph differences,so that the model can be transferred easily.Finally,integrate Ego-Transformer and WWGAN,design a unified framework for graph alignment based on transfer learning,and achieve intra-domain graph alignment and inter-domain difference elimination at the same time.Experimental results show that the proposed method improves Hits@k by 20.02%and MRR@k by 23.23%.(3)Aiming at the multi-domain graph alignment problem,this thesis proposes a domain generalization based multi-domain graph alignment method.Existing graph alignment methods only utilize single-domain data for model training,so there are certain limitations in robustness and generalizability.To solve above issues,this thesis aims to extract domaininvariant alignment patterns from multi-domain data,which will face the following challenges:how to extract domain-invariant features and how to extract domain-invariant alignment patterns.To address the above challenges,this thesis first proposes a novel domain-invariant feature extraction model,using a graph convolutional neural network as an encoder,and using a generative adversarial network to extract domaininvariant features.Then,a novel domain-invariant graph alignment model is proposed,and local and global loss functions are designed and jointly trained to learn domain-invariant alignment patterns.Finally,the specific training steps of the model are described to improve the reproducibility of the model.Experimental results show that the proposed method improves Hits@k by 14.01%and MRR@k by 10.63%.