| Graphs are a ubiquitous data structure that can flexibly build various models by utilizing the property of universal applicability,which has been more and more used in the fields of data analysis and relationship mining in recent years.Graph representation learning uses linear and nonlinear dimensionality reduction methods such as machine learning and deep learning to extract features from complex networks by embedding the network structure and node information into another vector space to obtain the lowdimensional feature vector representation.Such low dimensional representation contains the structural information of the graph,which can be used for downstream tasks such as node classification,link prediction and community detection.Due to the arbitrariness of graph structure,graph representation learning models need to be constructed from different perspectives,such as focusing on the method of node connection,or embedding node representation by manifold learning.Based on many graph representation learning methods,we study the community detection algorithm of nonnegative matrix and attention graph convolution neural network.Nonnegative matrix factorization has the advantage of high efficiency compared with spectral factorization,but the graph representation is easily affected by the randomness of gradient descent optimization algorithm and input matrix representation.On the other hand,due to the arbitrariness of graph structure,its dimensional consistency and node sampling affect the core mechanisms of representation learning such as convolution and attention,which increases the complexity of graph representation.A good graph representation learning model also needs to consider the size of graph data and the diversity of graph data and these problems bring many challenges to the ability,efficiency and applicability of the model.The research work of this thesis includes:1.Aiming at the problems of unsatisfactory representation of nodes and difficult fitting of local features in the community detection algorithm of nonnegative matrix factorization,we propose Nonnegative Matrix Factorization with Structure Extension(NMF-SE).Through structure expansion,the structural similarity of adjacent nodes is strengthened and the density of connections between nodes is increase,so as to improve the performance of nonnegative matrix factorization in community detection.Structure extension introduces the concept of message propagation from graph convolution and the process is to enable nodes to transfer their own structure to surrounding nodes in a certain proportion,so that adjacent nodes can get each other’s topology information,and improve node representations,which makes the nonnegative matrix factorization more suitable for community detection.The results of community detection on synthetic networks and real networks show that NMF-SE algorithm can better incorporate prior information and effectively improve the accuracy of community detection in complex networks.2.Aiming at the low computational efficiency of the attention mechanism on the graph and the complexity caused by sampling nodes,we propose information aggregation gated attention mechanism.This mechanism uses the nodes similarity and manifold distribution features to realize the attention sharing of neighborhood nodes under non-sampling conditions,greatly improves the efficiency of attention operation on the graph,and solves the problem of data redundancy and computational redundancy caused by sampling mode.The information aggregation gated attention mechanism uses the logistic regression function to calculate the attention between nodes,which optimizes the traditional attention operation mode on the graph and avoids the attention operation on additional sampling nodes,giving full play to the characteristics of attention locality representation and multihead-attention representation.3.We propose Information Aggregation Gated Attention Graph Neural Network(IAGAGNN).By parallelizing sparse matrix graph convolution and embedding neighborhood nodes with efficient aggregation function,information aggregation gated attention and graph convolution are organically integrated to significantly improve the model efficiency and model ability.IAGAGNN can dynamically assign different weights to nodes and neighborhood nodes for information propagation,improved the flexibility and representation ability of the model.Experiments show that the model achieves good performance in various tasks and various graph data sets,such as transductive task,inductive task and the latest graph data benchmark.Compared with the state of the art method,it improves the node prediction accuracy,and has the advantage of model universality.In addition,through the statistics of attention coefficient and the visualization of node community distribution,it provides the ability to further reveal the underlying relationship between nodes and communities. |
