Domain Adaptation Model Via Graph Convolutional Network

Transfer learning has attracted much attention since the beginning of machine learning research.It aims to make the algorithm obtain strong generalization abilities like humans done.Domain adaptation is one of the branches of transfer learning.Given two domains with strong similarities but different sample distributions,the domain adaptation problem requires training a classifier from samples and labels in the source domain,and to obtain well classification in the target domain samples which has no label or only a few of labels.Thanks to the powerful feature extracting performance of neural networks,many researchers have proposed a series of algorithms to deal with the problem of domain adaptation based on deep convolutional networks.However,these algorithms mainly focus on designing different distribution metrics and corresponding loss functions.The deep convolutional networks,which are based on the convolutional layer and the fully connected layer,do not pay attention to the relevant information between samples.When performing feature transformation by domain distribution loss,it would have a risk to transform related target domain samples into different categories,resulting a reduction of classification performance in the target domain.In recent years,researchers have successfully optimized the graph convolutional model to a network layer,which can be learn through back propagation similar to neural networks.Graph convolutional network has been applied to several transfer learning problems such as semi-supervised classification and one-shot learning.It uses the correlation graph between samples to fuse the relevant information between neighboring samples.It's effect in the domain adaptation problem is also worth to looking forward.In this paper,we first study a domain adaptation model based on graph convolutional networks.The algorithm uses pre-extracted features to construct a global correlation graph,and uses the graph convolutional network with maximum mean discrepancy loss to approximating the distribution of features between two domains.In this way,the model effectively retains the structural information between samples.The proposed algorithm achieved good classification results in some common domain adaptation data sets,proving the necessity of focusing on the correlation between samples in the domain adaptation problem.We further notice that the global graph used in graph convolutional networks requires all samples' features in advance.It resulted in that graph convolutional layers cannot directly fuse in deep convolutional networks for end-to-end learning,and the performance of graph convolutional networks depends on the quality of the pre-extracted features.We propose a trick that update the correlation map during training,and a sampling scheme based on labels and pseudo labels for minibatch learning.Through the combination of these two methods,the entire network can be learned by small batches of samples,further improving the performance of graph convolution network.