The Research On Heterogeneous Network Representation Learning Based On Multi-Channel Graph Convolution

In the information network,the basic method to determine whether there is a link between any nodes is to use the adjacency matrix.However,this method will have the following problems: if the node has only few neighbors,that is,there may not be connected edges between two random nodes,the adjacency matrix constructed based on these nodes will become sparse and consume a lot of unnecessary space for storage.Therefore,network representation learning needs to be introduced to deal with the above problems.Network representation learning aims to find a mapping function and establish a relationship between nodes and vectors,while preserving the topological properties of the network.Through this function,each node in the network is transformed into the corresponding low dimensional potential representation.Heterogeneous information networks describe the complex association relationship between different objects,and also include the semantics generated between different nodes and links.However,most of the existing research methods of network representation learning are based on the research of homogeneous information networks with the same type of nodes and link relationships,and ignore the diversity of node characteristics and link types in the network.At present,the representation learning of heterogeneous information networks is limited to the use of single-layer model when using two deep learning algorithms based on graph convolution neural network and graph attention mechanism,and does not fully capture the rich topology,node characteristics and semantic information of heterogeneous information networks.To address the above challenges,the specific work in this thesis is as follows:(1)A method for representation learning of Heterogeneous Dual Graph Convolutional Neural Networks,HDGCN,is proposed.construct feature maps and extract complex topology information in heterogeneous information networks according to different meta-paths.Secondly,a dual-channel graph convolutional neural network is used to obtain representations of network topology and feature attributes from different channels,and then combined with an attention mechanism,the semantic information and structural information generated by different meta-paths are fused for representation learning.Furthermore,by constructing a consistency loss function,we aim to fully capture the consistency of the two structures in the heterogeneous information network to enhance the feature commonality between different structures.(2)A method of representation learning based on Heterogeneous Adaptive Multi-Channel Graph Convolutional Neural Networks,HAN-GCN,is proposed.In order to further make up for the fact that HDGCN does not fully obtain the high-order neighbor information of the topology.Firstly,HAM-GCN extracts the complex topology information in the heterogeneous information network according to different meta-paths,and construct these topological structure information through the Positive Point-Wise Mutual Information(PPMI)matrix using a random process.Secondly,HAM-GCN simultaneously extracts specific and common embeddings about each meta-path from node features,topology and their combinations by designing three channels.Finally,combined with the attention mechanism,channel-level attention and semantic-level attention are designed to fuse low-dimensional representations obtained from different channels and different meta-paths.Channel-level attention focuses on learning the importance of each channel and assigns different attention values to it,and semantic-level attention aims to learn the importance of each meta-path and assign them appropriate weights.The collaborative optimization of the three channels is beneficial for embedding local consistency-based knowledge and global consistency-based knowledge,as well as learning some deep correlation information between topology and node features.Using two-level attention helps to select the most relevant channels and meta-paths.(3)This thesis adopts four real heterogeneous information network datasets DBLP,ACM,IMDB and YELP to conduct sufficient experiments,through three downstream tasks of node classification,link prediction and visualization,and parameter sensitivity analysis to verify HDGCN validity of the model.HAM-GCN performs three downstream tasks of node classification,link prediction and visualization,as well as parameter sensitivity analysis and ablation comparison Verify the effectiveness of the HDGCN model and the HAM-GCN model.Both HDGCN and HAM-GCN can outperform the compared baseline methods in node classification tasks,fully verifying that HAM-GCN has the ability to extract the most relevant informative node features,topology and semantic information generated by different meta-paths.