Research On Event-Semantic-Oriented Network Representation Learning In Heterogeneous Information Networks

With the continuous breakthrough of information technology and Internet technology,the network heterogeneous data containing the rich interaction relationships explode.These data describe different types of objects and their complex relationships,forming the ubiquitous heterogeneous information networks.The heterogeneous information network has the characteristics of structural nonlinearity,high-order interaction,information diversity,and dynamics,and contains abundant semantic information.Thus,it is difficult to represent them reasonably and efficiently in a unified way with the semantic information captured,which leads to the bottleneck in the efficient mining and application of heterogeneous information networks.In this thesis,we propose the research on event-semantic-oriented network representation learning in the heterogeneous information network,which aims to learn low-dimensional and dense vectorized representations of heterogeneous nodes,while preserving the event semantic information contained in the original network to the greatest extent,so as to break through the traditional research paradigm of network analysis based on the original adjacency matrix representation and support the analysis and application of heterogeneous information networks in the era of big data.This thesis is supported by the sub-topics of Beijing Municipal Education Commission co-construction project,i.e,"Research on information dissemination and evolution mechanism of heterogeneous information networks based on big data" and "Research on social-sensed cross-media data analysis and mining".This thesis focuses on the study of event-semantic-oriented network representation learning in heterogeneous information networks.Aiming at the four characteristics of heterogeneous information networks,including structural nonlinearity,high-order interaction,information diversity,and dynamics,the corresponding event-semantic-oriented network representation learning methods are proposed.The main contributions are as follows:(1)Aiming at the problem of learning the node vectorized representation which preserves the nonlinear network structure and captures event-based semantics in heterogeneous information networks,this thesis studies the network representation learning which preserves the nonlinearity of the structure,and proposes a novel method,namely Deep Extreme Learning Machine based Network Embedding(DELM-NE).This method constructs a depth model composed of multi-layer nonlinear functions to capture the nonlinear network structure,and designs an optimization function to capture the semantic information of events based on the first-order and second-order proximity between nodes.Experimental results show that this method achieves better performance than the traditional shallow models in network reconstruction and node classification tasks,and is 1.8 to 4.5 times faster than the traditional depth model in training running speed.(2)Aiming at the problem of learning the node vectorized representation which preserves the high-order interaction among multiple nodes and captures complex event semantics in heterogeneous information networks,this thesis studies the network representation learning which preserves the high-order interaction.On the one hand,a network representation learning method based on heterogeneous hypergraph,namely Event2vec,is proposed in order to capture the multi-level event-based semantics contained in intra-event relationships and inter-event relationships in heterogeneous information networks.On the other hand,a meta-graph-based network representation learning method,namely Meta-Graph based Heterogeneous Information Network Embedding(MGNE),is proposed to capture the multi-view event-based semantics in heterogeneous information networks.Experimental results show that the proposed methods in the study learn node representation vectors which can effectively support the classification,clustering and visualization tasks,and achieve better performance than the existing methods,which proves that by modeling the high-order interactions the methods can effectively capture the complex event-based semantic information from the network structure.(3)Aiming at the problem of fusing the original content information of nodes in heterogeneous information networks to learn the node representation that can further capture the event-based semantics,this thesis studies the network representation learning which fuses the node original content,and proposes a network representation learning method based on deep heterogeneous hypergraph embedding(DHHE).This method constructs the heterogeneous hypergraph,and designs a deep encoder for the original content which utilizes the deep learning to encode the visual content of nodes.Then the deep heterogeneous hypergraph embedding model is proposed to project different types of nodes into a common space,and the node representation and the encoded visual features are unified into one framework,which captures the event-based semantics underlying the network structure and the node content.Experimental results show that this method achieves improved performance on the application of image classification,group recommendation,and cross-modal retrieval,and the performance on image classification task is improved by at least 22%compared with the baseline methods.It proves that this method can effectively integrate the network structure and the original content of nodes,and learn the node representation that captures the event-based semantics.(4)Aiming at the problem of learning the node vectorized representation which adapts to the dynamic changes of the network and captures the semantic changes with the help of various information of heterogeneous information network,the thesis studies the dynamic network representation learning with multiple information fused,and proposes an inductive network representation learning method based on multi-task sequence to sequence learning(MINE).This method designs a new heterogeneous random walk method to capture sequence-based event semantic information,and converts the network representation learning problem into a one-to-many multi-task sequence-to-sequence learning problem to support inductive learning so that the vectorized representation of newly arrived nodes can be quickly inferred.Experimental results show that this method can improve the performance of node classification task by more than 5%compared with the baseline methods on the dynamic networks,which proves that this method can effectively generalize to the unseen nodes in the dynamic networks and learn the node vectorized representation that captures the event semantics.