Research And Implementation Of Node Representation Learning Algorithm Based On Graph Convolutional Neural Network

With the rapid development of artificial intelligence technology and the Internet industry,the amount of data in graph structure is exploding.How to fully analyze and apply these data has become a hot issue in academia and industry.Node representation learning is the generation of a low-dimensional,dense vector representation for each node in a network that enables the computer to process it effectively.The quality of node representation directly affects the effectiveness of downstream tasks.Therefore,how to integrate multi-dimensional network features in different scenarios to generate high-quality representation for nodes is a key problem to be solved urgently.The existing research shows that graph neural networks have excellent performance in node representation learning,but there are still some problems.In static scenes,existing methods ignore the unbalanced topology of nodes and fail to make full use of multiple information in the target neighborhood,which results in the representation of nodes being greatly affected by noise in the process of message propagation and low accuracy for downstream tasks.In dynamic scenes,existing methods make insufficient use of multiple information such as long-term and short-term sequential evolution characteristics of nodes,resulting in low accuracy of these methods when applied to downstream tasks.Moreover,most methods assume that the node set is fixed and only the edge set changes,resulting in limited application scenarios.For static attribute networks,a node representation algorithm NRLGAA based on graph convolutional neural network combined with active learning and attention mechanism is proposed in this thesis.In this algorithm,in order to avoid the unbalance of the distribution topology of labeled nodes,a method of constructing topologically balanced data set based on active learning is proposed.In order to extract and aggregate neighborhood features more reasonably,a neighbor node sampling and aggregation method based on graph convolutional neural network and attention mechanism is presented.Experiments on the public data set show that compared with the baseline algorithm,NRL-GAA significantly improves the accuracy of node classification.For dynamic attribute networks,a node representation algorithm NRL-GVW based on graph convolutional neural network combined with virtual nodes and sliding time window is proposed.In this algorithm,in order to make the model have the ability to process network snapshots of different sizes,a data preprocessing method of filling virtual nodes is proposed.For the joint extraction of the properties,structure and timing features of nodes,a modeling method based on graph convolution neural network and cyclic neural network is proposed,and a sliding time window is introduced to comprehensively consider the influence of long and short time series features on the current moment.Experiments on public data sets show that compared with the baseline algorithm,NRL-GVW significantly improves the accuracy of node classification and link prediction.Furthermore,a graphic structure data visualization prototype system is designed and realized for the node classification of citation network.The system integrates a variety of models,including NRL-GAA,and provides users with core functions such as graph upload,management and analysis application.