Research On Complex Heterogeneous Relation Modeling And Network Representation Learning

With the explosive growth of the Internet and information technology,complex and heterogeneous information is more and more ubiquitous.As an important data structure,a complex network can model and express the complex association information between different types of data objects.How to effectively learn the latent representation of the network structure data for different downstream network analysis tasks,such as node classification,link prediction,and anomaly detection,has become a hot topic in the field of machine learning and data mining recently.In the aspect of explicit relation modeling,the existing network representation learning methods do not systematically study the complex relation modeling problems of the interaction across multi-modal information,the modeling of high-order semantic relationship,and the fusion of multi-order heterogeneous relation;In the aspect of implicit relation modeling,how to mine the underlying association among highdimensional data is still an open problem.To solve the above problems,this paper explores the different types of data and complex heterogeneous relationship,and design different modeling methods,for the more powerful ability and better quality of network representation learning.Specifically,the main research contents and contributions of this paper are as follows:Firstly,we study the interaction across multi-modal information and representation learning methods on attributed networks.To address the problem of capturing the interaction across multi-modal information between network topologies and node attributes,we propose a general framework of representation learning on attributed network,which aims at jointly learning the node embedding and attribute embedding,to capture the interaction across multi-modal information between network topologies and node attributes.Moreover,we implement a dual autoencoder based anomaly detection model on attributed network(Anomaly DAE),in which,anomalies can be detected by measuring the reconstruction errors of nodes from both the structure and attribute perspectives.Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed method.Secondly,we study the modeling of high-order semantics relation and representation learning methods on heterogeneous information networks(HINs).To address the problem of traditional heterogeneous information network representation learning methods that neglect the high-order pairwise relation among nodes,we propose a representation learning framework named Heterogeneous Hypergraph Variational Autoencoder(Hete HG-VAE).It first maps a conventional HIN to a heterogeneous hypergraph with a certain kind of semantics to capture both the highorder semantics and complex relations among nodes.Then,deep latent representations of nodes and hyperedges are learned by a Bayesian deep generative framework from the heterogeneous hypergraph in an unsupervised manner.Moreover,a hyperedge attention module is designed to learn the importance of different types of nodes in each hyperedge,to improve the interpretability of the model.Extensive experiments on realworld datasets demonstrate the effectiveness and efficiency of the proposed method.Thirdly,we study the representation fusion and representation learning methods on multi-order relation networks.To address the problem of traditional network embedding methods that cannot learn the multi-order relation,in which,both the loworder pairwise and high-order complex relation are concurrent,we propose a general framework of High-order Relation Preserved Embedding(HRPE),which considers both low-order and high-order node relationships simultaneously.We use graph and hypergraph to model low-order pairwise and high-order complex node relationships respectively,and design a fusion module to fuse embeddings of different relation levels.Experimental results on real-world datasets for different tasks demonstrate that HRPE outperforms the state-of-the-art approaches,which proves the effectiveness of the proposed method.Lastly,we study the underlying correlation among high-dimensional data and its representation learning methods.To address the problem of implicit semantic relation modeling of high-dimensional data,we propose a network embedding based representation learning framework for high-dimensional data,which correlates the data into a form of a network structure according to their proximity in the feature space.During the feature learning,we introduce a dual-encoder module that consists of a feature encoder and a graph encoder to embed both the attribute of the sample itself and the relation among samples into the latent representation space.Moreover,we implement a correlation-aware unsupervised anomaly detection model,which fits both the distribution of attribute and relation for normal data,to detect out-of-distribution anomalies.Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed method.