Research On Unsupervised Domain Adaptation Method Based On Graph Convolutional Network

With the increasing cost of data labeling,the unsupervised domain adaptation task extracts relevant information from the source domain with label information to solve the target domain task without label information,which greatly reduces the cost of data labeling.In recent years,the popularity of graph convolutional networks has also led to great success in unsupervised domain adaptation tasks.Although previous unsupervised domain adaptation methods based on graph networks help the target domain to complete tasks by promoting inter-domain knowledge transfer,these methods have not yet explored and solved the following problems:(1)The problem of distribution differences between samples of the same class from different domains;(2)The target domain information in the domain adaptation task has not been fully mined;(3)There is a lack of effective extraction of inter-domain sample category information in domain adaptation tasks.Therefore,the following solutions to these problems are proposed in this paper:(1)Aiming at the problem of distribution differences between samples of the same class in different domains in the domain adaptation task,the Dual-aligned Unsupervised Domain Adaptation Method Based on Graph Convolutional Networks(DUDA-GCN)is proposed.The DUDA-GCN method trains a classifier adapted to the target domain task by aligning the distributions of samples in different domains and for same class in different domains.The framework consists of two parts,a cross-domain feature extraction module and a dual distribution alignment module.The former employs dual-channel graph convolutional networks with shared weights to learn common feature representations for both domains.The dual distribution alignment module includes an adversarial domain discriminant module,a pseudo-label generation module and a category alignment module.The purpose of the adversarial domain discriminant module is to maximize the domain discriminant loss and reduce the distribution difference between domains.The purpose of the category alignment module is to align the samples distribution in different domains of the same class.The pseudo-label generation module generates pseudo-labels for unlabeled samples by minimizing the pseudo-label generation loss.(2)Aiming at the problem that the sample information of the target domain is not fully utilized in the domain adaptation task,the Biased Target Domain Distribution Unsupervised Domain Adaptation Method Based on Graph Convolutional Network(BDUDA-GCN)is proposed.The BDUDA-GCN method reconstructs the target domain embedding feature representation to keep it similar to the original target domain sample distribution.The framework consists of two parts,a cross-domain feature extraction module and a dual distribution alignment module.The former uses two graph convolutional networks with shared weights to extract the latent feature representation of the two domains.In the cross-domain feature extraction module,in order to ensure that the target domain feature representation maintains the original target domain sample distribution,the target domain selfencoding module is designed to reconstruct the target domain feature representation at the feature level and the graph connection level.The dual distribution alignment module also includes an adversarial domain discriminant module and a category alignment module.The purpose of the adversarial domain discriminant module is to reduce the difference in domain distribution,and the purpose of the category alignment module is to align the distribution of samples of the same class in different domains.(3)Aiming at the problem that the relevant category information cannot be well extracted during the model training process,the Domain Segmentation Unsupervised Domain Adaptation Method Based on Graph Convolutional Networks(DSUDA-GCN)is proposed.The DSUDA-GCN method extracts the common category information of each domain by segmenting domain-specific and domain-common features for subsequent classification tasks.The framework consists of two parts,a domain segmentation module and a dual distribution alignment module.The former first utilizes a graph convolutional network to extract common latent feature representations for each domain.Secondly,the target domain self-encoding module is designed to reconstruct the source domain feature representation,and the reconstructed samples participate in training with the source domain and target domain samples.Finally,domain-specific feature representations are extracted from samples in each domain,and the specific and common feature representations extracted from each domain can be reconstructed into original samples to ensure complete information.The dual distribution alignment module is divided into an adversarial domain discriminant module and a category alignment module,respectively,to reduce the distribution difference between domains and the distribution difference between different domains of the same class.Three real paper citation network datasets(Citationv1,ACMv9,and DBLPv7)are selected for the above methods for experimental evaluation and analysis,and those methods compared with other excellent domain adaptation methods.In this paper,the experimental results show that all three unsupervised domain adaptation methods proposed are effective.