GAN-based Neural Networks For Graph Representation Learning

As a powerful data structure for modeling real-world problems,the graph has great practical significance to reveal the underlying structural characteristics of graphs.However,compared with structured data such as images,videos,and audios,graph data is more difficult to analyze effectively due to its high-dimensional and structured features.Graph representation learning aims to map high-dimensional unstructured graph data into a lowdimensional dense vector space and preserve graph structural features.Low-dimensional vector representations serve as a base to support various types of AI models for downstream tasks such as link prediction,node classification,and visualization.This study addresses the following questions:(1)Most generative adversarial network(GAN)-based research applies adversarial learning strategies directly to vector representations rather than deep embedding mechanisms,which do not take full advantage of GAN;(2)The proximity division between vertices is not detailed enough,which affects the performance of vector representation in downstream tasks;(3)Explore more flexible and effective ways to capture higher-order structures of graphs.The main work of the thesis includes:First,deep neural network GAN for graph representation learning(Dnn GAN).Unlike most generative adversarial models that apply adversarial strategies directly to the updating of vertex representations,this paper proposes to use the reconstruction mechanism of deep autoencoders as part of the adversarial learning structure,so that the adversarial learning strategies act on the deep embedding mechanism of graph representations.Dnn GAN consists of a generator and a discriminator,where the discriminator is a deep autoencoder,which can capture the higher order nonlinear structure of the graph,and the generator is introduced into the adversarial learning system as a competitor to form an adversarial embedding learning mechanism.A series of experimental results demonstrate the effectiveness of the model.Relevant work is being submitted(second trial,minor revision).Second,motif-aware adversarial graph representation learning.As the basic structure of graphs,Motif plays a crucial role in revealing the complex features of the graph structure.This paper proposes to use Motif to mine rich higher order structural information and combine it with the original lower order structure as the target structural condition.The generative adversarial model is constructed according to the target structural condition to generate the corresponding embedding representation.Experimental results on public datasets show that Motif GAN achieves substantial progress in various applications(link prediction,node classification and visualization).Relevant work has been published(SCI indexed).