Research On Multi-Relation Networks Embedding Based On Variational Auto-Encoders

Multi-relation networks contain the structured information of objects and various relations existing between objects,generally represented in the form of a graph.And the graph is composed of nodes representing objects and edges representing various relations.Compared to the single relation network,the multi-relation network can include more complex relations between objects.Therefore,the multi-relation network has a wider range of applications,such as biology,social science,linguistics,etc.However,most existed graph embedding methods tend to be applied to the single relation network rather than the multi-relation network.As the complexity of relation data increases,researchers are increasingly aware of the importance of multi-relation network embedding.Meanwhile some existing multi-relation network embedding methods often divide a multi-relation network into multiple single-relation networks in graph representation.Although this representation method can greatly retain the topology information,it also makes the size of input data too large,which leads to much problem such as too many model parameters and too high training calculation complexity.In response to the above problems,this paper attempts to introduce related technology of knowledge graph and graph signal processing.At first,we tried to introduce the multiple relations representation method in the knowledge graph for the first time into the multi-relation network embedding,combined with Variational AutoEncoders,and improved the loss function.Thus,we propose a multi-relation network embedding model PR-VAE.Because the graph representation method used by PR-VAE still loses some topology information,this paper attempts to retain more network information by dividing subgraphs.For this purpose,we design a new score function that can apply to the requirements of graph representation.We also use Variational Graph Auto-Encoders as the embedding model.Above all,we proposed a new embedding method MR-VGAE.Finally,because MR-VGAE has the bad performance on the embedding of “1-n”,”n-n”,”n-1” relations,we attempt to introduce the filters in graph signal processing to represent the multiple relations.Corresponding improvements have been made in graph representation methods and relational calculation processing,while retaining the MR-VGAE embedding framework.A new multi-relation network embedding model SD-VGAE based on spatial decomposition is proposed.Through the evaluation of experiments,we prove that our three methods are all superior to most baseline models in embedding effect and embedding stability.Multi-relation network embedding method proposed in recent years are mostl y unsatisfactory in terms of embedding effect and the difficulty of training due to problems in the design of graph representation method and the design of loss function.In this paper,we try to introduce new ideas to design some new graph representation methods,and propose some suitable score function.At the same time,combined with the actual problems encountered in the embedding process,the embedding model and loss function of the Variational Auto-Encoders are improved to better complete the embedding of the multi-relation network.