Research On Methods For Representation Learning Of Relational Graph And Multi-graph Fusion

Relational graph refers to the data,which has topology between objects.In general,objects in this kind of data are also known as nodes,where the topologies are called relationships.As computer science and the Internet developed rapidly,relational graphs with inherent connections are ubiquitous.For example,the friendships between users in social networks,the internal connections between research objects in physical systems,the interactive networks between proteins,the knowledge graphs in semantic networks and many other domain data.Abstracting objects into nodes and describing different connections between objects by relational graphs,can help people understand the inherent meaning of data and explore its potential value more conveniently.Because of existing a large amount of relational graph data,the application of analyzing and mining relational graphs achieves more and more attention and gradually becomes an international research hotspot.In real application scenarios,structures of relational graphs are usually much more complex where there may exist multiple graphs and the topologies cannot only be described by single relations.Specifically,the complexities are mainly reflected in the following two aspects: diversity of relationships and multiple graphs.Therefore,this dissertation takes the relational graphs with complex structures as research objects(including single relation and multi-relation),it is gradually carried out around the two problems,namely representation learning and multi-graph fusion,and the corresponding solutions are proposed.The main research contents and innovations of this dissertation include:(1)For attributed graph representation learning problem of single relational graph representation learning,this dissertation proposes a single relational attributed graph learning framework,namely TLVANE,which is based on a two-level auto-encoding.The framework leverages Doc2 Vec to learn the representations of textual attributes,and combines them with the adjacency vectors of nodes.Then,an end-to-end deep variational auto-encoding is used to learn the low-dimensional vectors of nodes.The framework has two levels of variational auto-encoding processes,where the first level of variational autoencoding process denotes the implicit vector that fuses textual attributes and topological structures,and the second level of variational auto-encoding process generates the implicit vectors of mentioned two features.Through encoding and decoding processes,TLVANE is able to learn the cross-modal characteristics of nodes while capturing the highly nonlinear structure of the data.Finally,the superiority and effectiveness of TLVANE are verified on three textual datasets.(2)For multi-relational graph convolutional network problem of multi-relational graph representation learning,this dissertation proposes a multi-relational graph convolutional framework,namely MR-GCN,which is based on generalized tensor product.By generalizing the three-dimensional tensor product,it proposes a multi-relational graph convolutional operator MR-GCO,which is also based on generalized tensor product.Specifically,given an adjacency tensor of an undirected multi-relational graph,the corresponding Laplace tensor is calculated.Then,the tensor transform eigenbasis is obtained by the tensor eigenvalue decomposition of the Laplacian tensor.Finally,by analogy to the traditional spectral graph convolutional process,the convolutional operation in spectral domain is conducted by using the tensor transform eigenbasis.Since the convolutional operation is directly defined on multi-relational graphs,it can dig up the inner relevance among the relations.Because both the tensor eigenvalue decomposition and the spectral domain convolutional process of multi-relational graphs are defined on the tensor product,and the calculation method of the tensor product is applicable to any unitary transform,the multi-relational graph convolutional operator has generalizability.Through comparison with advanced baseline algorithms,in four real-world datasets,the performance shows the effectiveness of proposed MR-GCN.(3)For node alignment across networks problem of multi-graph fusion of single relational graphs,this dissertation proposes a model for node alignment across networks,namely INAMA,which is based on intra and inter attention mechanisms.The model takes into account both topology structure and node attribute information and converts the node alignment across networks problem into a classification task.Firstly,in order to leverage the topology information efficiently,the model defines matched neighbors to replace the original topology to maintain the local network consistency.Secondly,in order to distinguish the influence of node information propagation in local topology and across network structure,the model introduces intra and inter attention mechanisms respectively.Then,using a deep neural network of binary classification,the problem of node alignment across network can be converted to a binary classification problem.Finally,through comparison with advanced baseline algorithms,in six real-world datasets,the performances verify the effectiveness of the proposed INAMA.(4)For entity alignment across knowledge graphs problem of multi-graph fusion of multi-relational graphs,this dissertation proposes a model for entity alignment across knowledge graphs,namely KAGNN,which is based on knowledge-aware attentional graph neural network.The model can contain knowledge facts that retain the original semantics,and make use of the complex network reasonably and effectively.Firstly,in order to effectively model knowledge facts,a knowledge-aware attention mechanism is proposed,which can automatically identify the importance of each knowledge fact.Then,in order to reasonably learn entity representations of complex relationship networks,a graph convolutional network with Highway Gates is adopted.The graph convolutional network with Highway Gates can capture higher-order neighbor information while reducing noise accumulation between hidden layers.Finally,through comparison with advanced baseline algorithms,in three real-world across knowledge graph datasets,the performances verify the effectiveness of the proposed KAGNN.