Research And Application Of Heterogeneous Network Node Representation Learning Algorith

Because the network structure can better model complex relational data in the real world,many network representation learning methods were proposed.They can well embed nodes into low dimensional vector space,and reserve the attribute information and the topology information of the network.These vectors can be directly applied to downstream tasks.in order to simplify the complexity of the problem,the early methods assumed that the real-world network had only one node type and edge type,namely,homogeneous network.However,these methods ignore the heterogeneity of real-world networks,which have many types of nodes and edges.These networks are called heterogeneous networks.Therefore,it is necessary to fully consider the properties of heterogeneous networks and design a representation learning model suitable for heterogeneous networks.This thesis investigates the current situation and shortcomings of heterogeneous network representation learning models,and finds that most of them need to manually design meta-path to extract the heterogeneous information of the network.The design of meta-path needs strong domain knowledge.Later,this thesis proposes two heterogeneous network representation learning models.The first is a representation learning model based on edge-aware neighborhood aggregation.It can extract the attribute information of nodes in different types of neighborhoods,and embeds the local structure information on different types of edges into the representation vector.The second is the representation learning model based on link preference.It considers the implicit preference between nodes,and measure the node preference based on probability and angle.This model can retain more abundant network structure information in the node representation vector.Finally,this thesis applies heterogeneous network representation learning model to fake news detection.,which illustrates the application value of the model through experiments.