Research On GCN-based Representation Learning Of Complex Network

Many scenes in the real world can be formalized as non-European network data,and more and more large-scale heterogeneous and dynamic networks appear.In the traditional network analysis research,the characteristics of nodes need to be extracted manually,which largely depends on network types,expert experiences and task types.In contrast,network representation learning can learn the effective representation of a network by automatically learning a mapping function,which can keep the rich information in the network to the maximum extent.However,there are many challenges in current network representation learning methods: 1)These methods are mainly designed for homogeneous networks(i.e.composed of only one type of nodes and edges),which are the simple versions of real heterogeneous networks.Research on heterogeneous network representation learning is still insufficient;2)Most methods assume that nodes and edges of networks will not change,but real networks are often dynamic;3)Many methods only use the structure information of the network,ignoring the fusion of the attribute information of the node.In response to the above problems,this paper systematically carried out the following research.This paper extends the traditional algorithm for homogeneous networks to heterogeneous networks and proposes a heterogeneous graph convolution(HIGCN)algorithm based on neighborhood influence.Firstly,the heterogeneous network is transformed into multi-relation network according to the meta path.Then,several HIGCN blocks are stacked.Each of them can aggregate attribute and structure information by considering different influence of node neighborhood,and then fuse the aggregated information under different semantic relationships(meta-paths)through 1 ~* 1 convolution.The experimental results show that the performance of three tasks(node classification,link prediction,visualization)performed by HIGCN on all six open datasets are better than that of other baseline methods.In order to extract the dynamic information in the network,this paper proposes the DVGAE algorithm,which improves GCN to get the time-series graph convolution.In the framework of Variational Auto Encoder,the time-series graph convolution is used to learn the network representation.Experimental results show that this method is superior to other baselines in link prediction task with two dynamic network datasets.