Modeling Heterogeneous Edges To Represent Networks With Graph Auto-encoder

With the increasing amount of information carried by network data,how to mine valuable information from its network structure has become a research hotspot in recent years.The network represents the structure of the learning network,and the low dimensional matrix preserves its structural features.However,there is not only one relationship type between nodes in real networks.How to mine the information of complex structure from heterogeneous edge networks is more challenging than homogeneous networks.The regularized graph auto encoder(RGAE)model is used to learn the network representation of heterogeneous edge networks.The model decomposes the heterogeneous edge network into multi view network according to its edge type,and uses the cooperation among views to supplement and learn a more robust low dimension network representation.In RGAE model,shared and private auto-encoders are designed to capture consistent information and unique information among multiple views.Graph auto-encoder has a deep neural network structure which combines the idea of graph neural networks(GCN).Compared with other shallow models,it can capture higher-order nonlinear structural information between nodes.In addition,two regular functions,similarity loss and difference loss,are designed for RGAE model to constrain the extraction and separation of consistent information and unique information respectively,so as to avoid information redundancy.Similarity loss,difference loss and reconstruction loss of balance optimization are used to optimize the model parameters in the process of back propagation learning.Finally,the consistent embedding and unique embedding among the views of learning output will jointly represent the structural information of heterogeneous edge network for downstream applications.The experimental results show that the learning effect of RGAE model on the vector representation of heterogeneous edge network is better than that of the existing scheme,and it is obviously improved in the experiments of node classification and connection prediction of different data sets.At the same time,the experiment also proves the importance of multi view cooperation and the rationality and necessity of two loss function design.