Network Alignment With Interpretability Using Graph Neural Network

With the rapid development of social media,more and more users are involved in multiple social network platforms for communication,working,learning and entertainment activities.How to find common users in different social networks has become a key concern for many valuable applications.This type of problem is referred to network alignment,and the common users here are called the anchor nodes.Because it can be widely used in many fields,such as network fusion,link prediction,cross-network recommendation,etc.,the research purpose of this thesis is to address the network alignment problem in social networks.Existing methods generally consist of two parts.The first one is the network representation learning process,which aims at learning the representation of each user with a feature vector leveraging the network structure and the user's feature information.In this part,various graph/node embedding techniques are usually used.The second part is the matching process.That is,it learns a matching function to link anchor nodes between multiple networks based on the learned node representation vectors.However,these methods still suffer from the following problems.1)matching confusion: During the process of network representation learning,neighbor user nodes are tightly embedded in a single network,which makes it difficult to distinguish them between neighbor nodes,and,most importantly,we need to discriminate the true corresponding anchor nodes from their close neighborhoods during the matching process.In another word,the objectives of the two parts are contradictory,resulting in unsatisfactory network alignment effects;2)lack of uncertainty: the previous model represent nodes as simple points vector and ignore the potential uncertainty of node representation;3)Point-to-Point(P2P)matching constraints: previous methods use point-to-point to match anchor nodes,which would cause severe overfitting and affect the accuracy of network alignment;4)lacks of model interpretability: existing methods,especially deep learning based models,regarded the model as a black box and did not analyze the interpretability of the network alignment process.For example,which training data is most effective for successfully aligning a particular pair of anchor nodes? what are the real factors that affect the performance of network alignment?In this thesis,we propose two methods dNAME and UANA to solve the above problems.In particular dNAME(disentangled Network Alignment with Matching Explainability)uses graph neural networks and kernel methods to learn the representation of nodes,and then uses a matching function to link anchor nodes.Moreover,dNAME can effectively distinguish anchor nodes from their neighbor nodes,thereby solving the problem of matching confusion.UANA(Uncertainty-Aware Network Alignment),while being another graph neural networks-based model,utilizes Gaussian embedding to capture the uncertainty in the node representation learning and uses generative adversarial matching to solve the point-to-point(P2P)matching constraint issue.In addition,we also analyze the interpretability of the two models based on influence functions,which allows us to better understand the underlying mechanism and advantages/disadvantages of network alignment models.Finally,we validate the proposed methods on real-world datasets,and the experimental results show that the proposed two models outperform existing state-of-the-art baselines.