Research On Technique And Application Of Heterogeneous Graph Embedding

With the rapid development of the Internet,large amount of data is generated every day.So how to organize the data has become a very critical issue.In fact,most of the data in life can be described through the network,or in other words,graphs.For example,people's social relationships can be constructed as a social network with people as nodes and people's connections as edges.Based on the graphs,graph embedding technique is gradually emerging.Graph embedding is essentially a method of representation learning,which maps a graph into a low-dimensional space while retaining the structural informa-tion of the graph and the attributes of the nodes themselves.This technique benefits many downstream tasks,such as node classification,link prediction,recommendation systems,and so on.This thesis focuses on graph embedding methods based on deep learning.Starting from the earliest random walk-based models,various deep learning based graph embed-ding models have been proposed.Recently Graph Convolutional Network(GCN)per-forms well and attracted people's attention.It uses convolutional layers to aggregate neighbor nodes around the nodes to learn local neighbor information.However,this model uses single nodes as input,and can only learn limited local information.In addition,the definition of neighbors is solid.To solve the above problems,this thesis proposes a novel subgraph convolutional network,which uses subgraphs instead of single nodes as the input of the model.First,heterogeneous graphs are constructed based on existing data,in which entities are re-garded as nodes,and connections between entities as edges.Then subgraphs are extracted from the entire graph,and subgraph embeddings,i.e.,representation of the subgraphs,are learned as input of the subgraph convolutional network.Finally,subgraph embeddings are aggregated and node embedding are learned.This thesis mainly uses the recommen-dation system as the application scenario.The recommendation system can be regarded as a problem of link prediction,i.e.,predicting whether there will be a link(click)between the user and the item.The main content of this thesis is as follows:1)First the current development status of graph embedding technique and recommen-dation system is investigated,and a graph-based recommendation problem is formulated,in which users and items and their connections are represented in heterogeneous graphs.2)Based on the necessity of extracting subgraphs in this paper,two subgraph extrac-tion methods are proposed according to the application scenario in this thesis,and their advantages and disadvantages are analyzed.3)In order to solve the problems in existing graph convolutional network,a novel subgraph convolutional network is proposed.Subgraphs are extracted from the entire huge graph,and subgraph embeddings are learned with methods in natural language process-ing.The subgraph is more flexible compared to all first-order and second-order neighbor nodes.Because the input is a subgraph,which already contains relatively rich neighbor information,the number of convolution layers can be reduced,which is beneficial to im-prove the efficiency.4)This thesis uses a simultaneous learning strategy.In other words,graph embedding and downstream link prediction task are learned at the same time,which can improve the performance of the task in a targeted manner.Finally,extensive experiments are conducted on real-world data sets,including com-parison with other models and comparison of different parameters of the model itself.Through comparative experiments,we can see that the model proposed in this thesis per-forms better.