Research On Graph Generation Model Based On Adversarial Learning Method

With the development of social networks,recommendation systems,bioinformatics and other fields,graph data has become an important data type and more and more common.The task of graph generation has made significant progress in the past few years,and has attracted widespread attention from academia and industry.It refers to the generation of graphs with specific structure and attributes,such as molecular structure,circuit diagram,social network,etc.Among them,the Generative Adversarial Network(GAN)using adversarial learning is one of the most popular methods for generating deep learning graphics.However,there are also some challenges in applying adversarial learning to graphic generation tasks.For example,algorithms are prone to pattern collapse,generate samples with insufficient diversity,and are unable to generate graphics with highly complex structures and control the attributes of generated graphics.Based on the above problems,this paper first proposes an improved conditional adversarial graph generation models with residual connections(Graph-resCGAN).This model takes adversarial learning as the basic generative model to build discriminators and generators.Aiming at complex graph data,the model makes the following improvements:(1)Propose residual connection to deepen network layer to capture more graph information.For graphic structures that cannot capture information in depth,using improved residual blocks instead of traditional convolutional blocks can better capture the information of graphic nodes and edges.(2)Propose an improved conditional classification algorithm that better guides graph generation by constructing a hierarchical classifier and introducing graph label information as conditions.The improved conditional discriminator part distinguishes between real data and generated data category labels,enabling discriminators with classifiers to have the perception ability of generators.The final Graph-resCGAN model was validated for improvement effectiveness through design experiments.Compared with classical graph deep learning methods,this model performs better in generating data distribution performance MMD scores,with an average improvement of 8.74%on various datasets.On the basis of the Graph-resCGAN model,in order to solve the limitation of requiring predefined maximum nodes,and the universality problem,a conditional graph generation model based on variational generative adversarial network with attention mechanism(Graph-attVGAN)is proposed.The specific improvement of this model is as follows:(1)Combining variational autoencoder with adversarial learning.This model first embeds graphs of different sizes into a unified graph using a variational autoencoder,and then uses adversarial learning to obtain reconstruction losses and conditional discriminator losses.(2)Introduce graph attention mechanism.Propose combining graph convolutional networks with graph attention mechanisms to enable the model to learn vertex information more fully and obtain richer features.In the final experiment,it was proven that the Graph-attVGAN can achieve a satisfactory level in the evaluation indicators of results.The novelty increased by 2.64%and the uniqueness increased by 2.85%.Moreover,the model has better universality.