Graph Convolutional Network-based Multi-source Transfer Learning

The development of mobile internet has brought a lot of data with graph structure,such as social network and so on.More people pay attention to the data with graph structure.To deeply mine spatial structure information,graph convolution network is proposed.However,the graph convolution network can well mine the information of graph structure,when a new domain appears,the trained neural network can not be directly applied to the new domain without modification.Transfer learning can help us to solve the problem of new domain named target domain by using the knowledge of existing domain named source domain.It can be divided into single source and multi-source domain transfer learning.Actually,single source domain transfer learning is a special case of multi-source domain transfer learning,this study focuses on the situation of multi-source domain transfer learning.Two graph convolution network-based transfer learning approaches,graph Multi Source Tr LDA and Hegraph Multi Source TL are proposed to solve the as above problem as above of homogeneous social network graph as above and heterogeneous social network graph models,respectively.To solve the problem of transfer under the homogeneous situation,we propose a method of measuring the distribution distance based on the mixed measure,and select the source domains whose distribution distance are smaller to fuse,then align the distribution of the source domain and the target domain and finally use the graph convolution network to mine the spatial structure information and predict the label of target data.We test our multi-source homogeneous transfer learning method on open datasets with graph structure,and prove the effectiveness of our method.As for the heterogeneous social network graph model,we transfer the source domain data by using a small number of labeled target domain data to solve the problem that graph convolution network can not deal with the heterogeneous data,and then consider the heterogeneity between the source domains,we train a multi group graph convolution network,and use the stronger representation of graph convolution network ability to mine the shared information of different source and target domains,and finally carry out weighted fusion.We tested our transfer algorithm on OSN dataset,and compared with model without transfer,the result prove the effectiveness of our method.