| Deep learning models rely on a large amount of labeled data for training.However,the real-world data usually lack labels,and the cost of labeling data is unacceptable,and the multi-source data is difficult to satisfy the basic assumption of independent and identical distribution.These two reasons make it difficult for deep learning models to be deployed in real-world scenarios.Domain adaptation is a solution to this kind of problem.It mainly uses the existing large amount of source domain labeled data to train a model,so that it can be adapted to the target domain with little or no labeled data,and effectively solve the above-mentioned problems.Recent years have witnessed tremendous interest in deep learning on graph-structured data.However,current graph domain adaptation methods are generally adopted from tradi-tional domain adaptation tasks.Most of the existing work in this field is based on generative adversarial networks or gradient reversal layers to learn the domain-invariant representation of data.However,the generation process of graph data is highly uncertain and its domain-specific structure is highly complex,Existing methods generally ignore the properties of graph data,making the extraction of domain-invariant representations very difficult for graph data.There-fore,domain adaptation for graph-structured data is an equally important and challenging task.In this thesis,we study the domain adaptation problem for graph data and apply our proposed framework to graph classification tasks and cross-domain recommendation tasks,respectively.The specific work is as follows:（1）We propose a disentanglement-based unsupervised graph domain adaptation method and apply it to the graph classification task.Specifically,based on these properties of graph-structured data,we first assume that the generation process of graph-structured data is con-trolled by three independent latent variables,namely semantic latent variables,domain latent variables,and random latent variables.Semantic latent variables are controlled by label-related information,domain latent variables are controlled by domain-specific information,and ran-dom latent variables are controlled by the uncertainty and noise of graph data.Based on the above assumptions,we obtain a causal model of the graph-structured data generation pro-cess,and propose a disentanglement-based unsupervised domain adaptation method for graph-structured data.We first use a multi-layer graph convolutional network to extract node features.Then,we use graph variational auto-encoders to reconstruct these latent variables and disen-tangle them through three supervised learning modules to extract domain-invariant semantics information,which improves the performance of downstream graph classification tasks.In or-der to solve the uncertainty of graph data,we also design a graph data augmentation module,which simulates the uncertainty of the graph generation process by randomly adding or remov-ing edges,and use the synthesized noise as a supervision signal.The disentangling network of random latent variables is independently trained,which effectively improves the robustness of our model.To verify the correctness and effectiveness of the proposed model,we conduct a series of experiments on real-world datasets.The experimental results show that,compared with the existing domain adaptation methods,the proposed disentanglement-based graph do-main adaptation method can better extract the domain-invariant semantic information,so that the downstream graph classification task can achieve a higher1-score.（2）We further extend the domain adaptation framework for graph classification tasks to cross-domain recommendation tasks,and propose a disentanglement-based graph cross-domain recommendation method.Specifically,the semantic information in the graph genera-tion process is changed from the label of the graph sample to the interaction records between users and the items.Therefore,we learn the user node embedding from multiple user-item bipartite graphs and utilize a disentangling module to extract the cross-domain user interest representation.In order to adapt to the large amount of user and item nodes,we improve the graph convolutional network so that it only aggregates local node representations.We also use different weights to learn the asymmetric relationship between users and items,which im-proves the computational efficiency and expressiveness of the graph convolution layer.In or-der to verify the effectiveness of the proposed algorithm,we conduct extensive experiments on multiple real-world datasets.The experimental results show that,compared with the existing cross-domain recommendation methods,The proposed disentanglement-based graph cross-domain recommendation method has an average improvement of 14.11%on HR@5 and an average improvement of 15.32%on NDCG@5.It shows that our proposed method can ef-fectively disentangle the shared user interest representation from the structural information of the multi-domain user-item bipartite graph,and provide more accurate recommendations for cold-start users in the target domain.Finally,the experimental results also verify that the mod-eling of the graph generation process can effectively solve the entanglement of domain-specific structural information and domain-invariant semantic information. |
