Node Embedding Representation Based On Graph Convolutional Neural Networks

In the real world,there are many graph-structured datasets,such as social graphs,citation graphs,etc.These graphs can also be divided into homogeneous graphs and heterogeneous graphs.Homogeneous graphs have only one node type,while there may be two or more types of nodes in heterogeneous graphs.Analyzing such datasets can reveal a lot of hidden information,which has become a hot topic in academia and industry recently.However,these datasets cannot be directly inputted into existing algorithms.So an urgent problem to be solved is how to represent the graph and retain the information of the original graphs.In recent years,network representation learning aims to represent a node with a low-dimensional vector,that is,two nodes directly or indirectly connected in the original graphs should also be close to each other in the low-dimensional space.This paper focuses on the graph embedded representation both in homogeneous and heterogeneous networks.The contents are as follows:(1)In chapter 3,our method is based on graph convolutional neural network(GCN).The label is used as the anchor points of different node types in embedding space.In the optimization process,the penalty function of distance between anchor points and node vectors is introduced,which can solve the problem of overfitting in traditional GCN.At the same time,the relevance between labels are considered.the experiment shows that our method can improve the classify accuracy in multi-label graphs.(2)In chapter 4,The attention mechanism of feature dimension is introduced.We use a matrix to represent node features instead of a row vector.For the feature matrix of a node,a multi-level graph attention network is proposed,which can allocate different attention weights for the neighbors of a node.Finally,we execute our model in three graph datasets.Experiments show that our model is superior to the most advanced method.(3)In Chapter 5,a new embedding method in heterogeneous networks is proposed.The model represents a node with its neighbor nodes by several convolutions and pooling operations,and takes the final vector as the vector representation.At the same time,different loss functions are used to optimize the vector representation of nodes.The experimental results in several datasets show that the model can achieve the state-of-the-art results in link prediction and recommendation tasks.