| Non-grid data widely exists in our life,the analysis and processing of non-grid data play an important role in the further application.In recent years,graph neural network has attracted extensive attention,especially graph convolutional network.At present,the work of graph convolution network mainly focuses on the network architecture and feature propagation.This thesis focus whether the loss function that can approximate the global information of the graph can improve the performance of the model.The main work is as follows:(1)From the perspective of data distribution,this thesis assumes that the data distribution of citation network data obeys the distribution of each sample around its class,and introduces the center loss function.Center loss can measure the distribution information of data,and constraint model learning more discriminative features.In this thesis,the center loss function is introduced into the graph attention network with cross-entropy loss to form a joint loss function.The experiments of node classification on Cora,Citeseer and PubMed datasets show that the result has a certain improvement.(2)The optimization method based on edge reconstruction in network representation learning can retain the main structure information of data during the training process.From the perspective of data structure information,this thesis learns from the optimization method of maximizing reconstruction probability,and introduces a joint loss function composed of first-order edge prediction loss and cross entropy loss.Edge prediction loss represents the structural loss of dataset and the global information of graph from the side.The experiments of node classification on Cora,Citeseer and PubMed datasets is tested and the results were improved to some extent.(3)As an unsupervised learning method,auto-encoder is used to learn representative features in network representation learning.This thesis adds an decoder containing two layers of attention layer after the last layer of the graph attention network.The reconstruction loss between output features and input features was introduced as well.The model is constrainted to learn features that can represent the graph global information.Node classification results on Cora and Citeseer datasets are also improved. |
