Research Of Seed Set Expansion Algorithms Based On Graph Neural Network

Graph data exists widely in the real world.Seed set expansion is an important task for graph data analysis which plays a key role in the implementation of recommender systems,social network analysis,biomedical and other applications.Given a node set or subgraph set as seeds,the seed set expansion problem aims to obtain more nodes or subgraphs similar to the seeds through the expansion algorithms.Algorithms proposed by existing related research work often only use graph topology information to expand,which ignore node attributes information.At the same time,most works only focused on the expansion of node sets,while the problems of subgraph set expansion are rarely studied.Therefore,this dissertation focuses on the problems related to seed set expansion,and research conducts on node set expansion algorithm and subgraph expansion algorithm,designs and implements a music recommendation prototype system based on the expansion algorithm.The main research contents and results of this dissertation include:(1)A Node Set Expansion Based on Graph Neural Network(NEGNN)is proposed.Aiming at the shortcomings of traditional algorithms that only use topology infomation of graph to expand,this paper uses graph neural network to learn node features and predict node labels,and uses Expectation-Maximization algorithm to alternately train two Graph Attention Networks models with predicted labels for iterative optimization.In the process of model training,the attention coefficients about node attributes and label propagation are learned,finally the attention coefficients are used to guide PageRank algorithm to walk,and the expanded node set is selected according to the walking results.This paper conducts experimental evaluation on three datasets of PubMed,Citeseer,and Cora,and compares it with the node set extension baseline algorithms.The experimental results show that the NEGNN method is 5.09%higher than the baseline algorithms on average under the recall index.(2)A Subgraph Set Expansion Based on Graph Neural Network(SEGNN)is proposed.Aiming at the problem of subgraph set expansion,this paper uses the Graph Isomorphism Network(GIN)model to learn node features,uses max pooling and context-aware attention to obtain subgraph features information,and then uses the Neural Tensor Network(NTN)model to learn the similarity between subgraphs and neighbor subgraphs,and the similarity are used to guide the PageRank algorithm to walk,and obtains the extended subgraph set according to the results of the walk.This paper uses six datasets of PTC-MR,DD,ENZYMES,PROTEINS,NCI1,and MUTAG for evaluation.The experimental results show that,the SEGNN algorithm proposed in this paper is better than the unsupervised classification method Infograph and Graph Transformer Networks(GTN)under the recall index,2.83%higher on average.(3)The song features are extracted according to the information such as the singer and the album to which the song belongs,and the songs are regarded as nodes.The edge of graph data is added to the songs,the songs collected by the user are used as seeds,and the music recommendation problem is transformed into a node set expansion problem.The paper uses NEGNN as the recommendation model,and combines this model and adopts the SSM framework to realize a music recommendation prototype system.