Research On Link Prediction And Graph Label Prediction Algorithm Based On Network Structure Feature Representation

In recent years,there have been a large number of studies on complex networks,which has become one of the research hotspots in the fields of computer,biology,sociology and so on.Complex network structure can be used to describe all kinds of systems with high technology and intellectual importance,among which the structure of complex system and the relationship between system structure and system function are very important concerns.The main research on complex networks is based on the theory and method of graph theory.Among the various research tasks,link prediction and graph classification are very important research hotspots.However,in the traditional link prediction task,the algorithm based on community structure is often limited by the community division algorithm.When the number of community connections is very small,the prediction accuracy of this algorithm will be very low.In the graph classification task,network noise is inevitable,and large-scale network will increase the time of model training.At the same time,in the graph classification task,the fixed neighborhood relationship leads to some important node features can not be effectively expressed,which impedes the development of the predictive graph label task.In view of the above problems,the main work and innovation completed in this paper are as follows:(1)A multi-resolution link prediction method based on community correlation is proposed.Based on community similarity,this method is more suitable for predicting the probability of missing link between node pairs with a long distance.By dividing multi-resolution communities and combining them with the connections between communities,the probability of the existence of link can be well predicted even when the number of community connections is small.In the final experiment,two network data sets of different scales are used to compare the prediction accuracy with other common algorithms.The experimental results show that the proposed method has excellent competitiveness.(2)A graph label prediction method based on local feature representation is proposed.By extracting network substructure information,this method can greatly shorten the running time of the prediction algorithm and reduce the influence of noise on the prediction results to a certain extent.At the same time,the physical meaning of the between centrality node is fully considered,and the substructure set is obtained through depth-first search.This set of substructures not only contains the characteristics of each region in the molecular graph,but also well preserves the connections between each region.In addition,this method defines a new distance index,divides the substructure set into similar set and dissimilar set,and then combines the characteristics of similar set and dissimilar set to train the model.Experiments show that the performance of this method is better than other methods on different datasets.(3)A method of graph label prediction based on attention mechanism is proposed.By using the attention mechanism,this method changes the inherent relationship between nodes and neighbourhood nodes,and can specify the connection weight between nodes and their neighbors.This mechanism changes the connection weight according to the eigenvectors of nodes and their neighbors,and amplifies the role of important node eigenvectors in the network.At the same time,combined with the traditional GCN network,more distinguishable graph vector can be obtained.This method also discusses the effect of different Readout methods on the precision of the predicted graph label after using the attention mechanism.Finally,by comparing with other methods,it is proved that the proposed method has better ability of predicting graph labels.