Research And Implementation Of Graph Representation Learning Algorithm Based On Network Sampling And Feature Coding

Representation learning can encode the information of graph-structured data into the low dimensional vector space,which makes it more convenient to process and calculate graph-structured data in computer than before.It is important not only to reason and calculate the knowledge in graph-structured data,but also to promote the application of the knowledge graph in the development of social intelligence.By summarizing the current researches in the field of representation learning,this thesis proposes a graph representation learning algorithm framework based on network sampling and feature coding.At the same time,it is found that the current algorithms have some short comings: First,it is easy to ignore nodes' different features in different local network structures by taking the entire network as the algorithms' input,which also makes it difficult to distinguish some adjacent nodes when dealing with complex graph-structured data.Second,the model based on uniform sampling of nodes and attributes cannot effectively distinguish the importance of different attributes on the association between nodes,which limited network information that contained in each single nodes' sampling sequence.In order to deal with above problems,this thesis puts forward corresponding solutions from two aspects: multi-network analysis and weighted attributes sampling.And then studying how to design an effective feature coding model to extract the features in the sampling sequence.The main innovations and contributions are as follows:(1)A representation learning algorithm based on multi-network analysis is proposed.Based on the mixed sampling of nodes and attributes,the algorithm can obtain sampling sequences containing different subnet results by setting different hops.The experimental result shows that the algorithm can effectively distinguish the node characteristics under different local network structures,and the algorithm's performance of node classification on three public datasets are significantly better than relevant works,and the overall average accuracy performance is improved by about 12.59%.(2)A representation learning algorithm based on mixed sampling of nodes and weighted attributes is proposed.On the one hand,the algorithm calculates the attributes' sampling weight through the distribution of nodes' common attributes,and then obtains the sampling sequence through the mixed sampling of nodes and weighted attributes.The experimental result shows that the mixed sampling method can effectively distinguish the contribution of different attributes to the association between nodes and increase the sampling probability of nodes with the same label in the sampling sequence,which make the sampling sequences contain more effective information.On the other hand,a feature coding model combining bidirectional long-short term memory and attention mechanism is proposed to extract the network structure and semantic features in the sampling sequence.The experimental result shows that the node classification's average accuracy of the representation learning algorithm on three public data sets is82.41%,which is 1.20% higher than related works.Based on the above work,an application verification system of representation learning algorithm is developed,which has visualized the comparison and analysis of the application effects of different representation learning algorithms on graphstructured data.The practical application shows that the proposed algorithms can alleviate the problems of nodes' feature extraction in large-scale networks and sparse attributed-networks partly,and improve the vector representation quality of graph-structured data and the performance on downstream tasks.