Subspace-based Node Representation Learning On Networks

In the real world,a large amount of data is often expressed in the form of network,which can effectively depict the relationship between individuals.Network is also one of the most intuitive and effective data modeling tools for real complex systems.Therefore,in recent years,network mining has gained great attention in various fields and has been widely used in many practical problems.Representation learning,as one of the key tasks in network mining,aims to learn node representations in low-dimensional space,so as to help downstream tasks to further explore potential patterns in networks.At present,there are many methods of node representation learning to be proposed.Such methods still face many challenges.Firstly,due to the complexity of network,traditional methods cannot learn node representations with obvious differentiation and thus fail to identify different connection modes of network.Secondly,more networks not only contain complex topology structure,but also high-dimensional node attributes in the real world.This leads to inconsistency problems in attribute networks,which makes traditional representation methods based on global homophily cannot effectively integrate network topology and node attributes.In addition,due to attributed networks carrying noisy information in the real world,how to learn robust node representation is also a difficult problem in current research.For the above challenges,this dissertation proposes a series of new methods,according to the complex,noisy and heterogeneous characteristics for learning node representations by exploring the subspace structure of networks.The main research contents and contributions of this dissertation are as follows:1.For the problem of weak ability of node representation differentiation caused by the complexity of network topology,this dissertation proposes a subspace-based and cluster-driven method to learn node representation for obtaining different connection modes of nodes.This method aims to learn common and distinct patterns of network connections by non-negative matrix decomposition.In order to further distinguish the learning network connection features,the two patterns are constrained with given label information,so that the network patterns of different categories are as different as possible,while the network patterns of the same category are as same as possible.Such constraints can embed distinct patterns into the node representation.To verify the effectiveness of the algorithm,the learning node representation with the distinct patterns is utilized to the diagnosis of patients with schizophrenia and major depression.Experimental results show that the accuracy of the proposed method is more than 15% higher than that of the existing methods.2.For the problem of inconsistency between topology and high-dimensional attribute information in attribute networks,a structured subspace representation method on attribute networks is proposed in this dissertation,which is used to learn the node representation integrated by topology information and node attributes in different subspaces.This method uses latent features as a bridge to represent network topology and node attribute information,respectively.Furthermore,constraints are designed on topology representation and attribute representation to obtain community structure and learn consistent representation in each subspace.Experimental analysis on extensive real-world data sets shows that the node representation learned by this method is not only superior to state-of-the-art methods in community detection and node classification,but also interpretable.3.For the problem of different patterns of community structure in attributed networks due to different generating mechanisms,this dissertation proposes a node representation model based on community attention network to learn the node representation of exclusive community patterns.By learning the corresponding mappings of different communities,this model learns the exclusive patterns of different communities.The center of each subspace cluster is determined by using some labeled nodes as anchor points combined with the network connection patterns of unlabeled nodes.The model uses the exclusive constraints of community features to strengthen the relationship between labeled nodes and unlabeled nodes.This model then learns the mapping of different community patterns and obtain node representation with the patterns of communities.Extensive experiments show that this model can outperform some state-of-the-art graph convolution models when given fewer labelled nodes on real data sets.4.For attributed networks with noise,this dissertation proposes a subspace-based and robust method for learning node representation.The method firstly transforms node attributes into edge attributes to eliminate the noise attributes,and further such edge attributes and topology are represented by tensors for representing attributed networks.To learn the network community structure,this method introduces community-level constraints and attribute selection constraints and then learns the topology and attribute information related to the community.Such constraints ensure the method to eliminate irrelevant attributes and learn robust node representations.Experimental analysis on synthetic and real world datasets show that this method is superior not only to linear methods in community detection,link prediction and node classification,but also to some nonlinear methods.5.For over-smooth problem of graph neural networks,this dissertation propose a node representations model of community-driven graph convolution network,which aims at learning node representations relying on different communities.This model limits the propagation of node attributes inside of community structure by detecting communities,which avoids the model over-smooth.In particular,by learning a weight matrix,the model deletes edges between communities to ensure that node representation do not propagate information across communities.At the same time,the model uses modularity constraint to learn high-quality community structure,and finally learns node representations with community patterns.Experiments show that this model can effectively remove the noise edges between network communities,and learn robust network node representation.The model is 4% better than the classical model on Cora,and has strong robustness to the introduced parameters.