Network Structure Representation Based On Complex Network Characteristics

At present,more and more data presenting complex network structure have been discovered,such as social network data,postal network data,molecular network,etc.These complex network data contain rich information,so a large number of studies on complex network have emerged in recent years.Community evolution prediction and graph classification are important research problems in complex networks.The selection of timeframes division mechanism in the prediction of community evolution events has an important influence on the whole prediction model.Most of the existing prediction models adopt fixed timeframes division mechanism,such as disjoint timeframes division mechanism and overlapping timeframes division mechanism.However,due to the fixed setting of window size and overlap rate,this kind of method is not suitable for real networks with strong variability,that is,this kind of mechanism is not universal.In the problem of graph classification,most models use the features of the whole graph to learn,which will increase the training time of the model.The noise and redundant information in the graph will also affect the prediction result of the model.In addition,there are also models that use the local structure features of graph to learn.But these models use fixed neighborhood relations in graphs,which will lead to some important nodes cannot be effectively expressed,thus affecting the accuracy of classification.In order to solve the above problems,the main work and innovation of this paper are as follows:(1)A new community evolution tracking model is proposed.Based on the universality of timeframes division framework,an optimized Self-Adaptive timeframes division mechanism is proposed.This mechanism can adaptively adjust the size and number of timeframes according to the activity of the network.This method can get the most suitable network partition.In addition,a community evolution tracking model is established based on the SelfAdaptive timeframes division mechanism.Through experiments on several real networks,this method can reduce the loss of information in the process of tracking communities in different periods,and improve the quality of the community evolution sequence.(2)A new community evolution prediction model is proposed.This method optimizes the tracking quality of community evolution events under the Self-Adaptive timeframes division mechanism.In addition,more features are proposed to describe the state of communities.Then,evolution events of network communities are predicted by community evolution prediction model.The real network experiment shows that our community evolution prediction model can achieve better prediction accuracy.(3)A new method for substructure partition of molecular graph is proposed.This method uses the meaning of the topology features of intermediary centrality and combines depth-first search to obtain substructure.The substructure extracted by this method can minimize the influence of noise and greatly shorten the running time of the algorithm.In addition,a new network structure characterization method is proposed for the substructures.This method makes the representation of network structure feature more comprehensive.Through experiments on different data sets,this method has better performance in graph classification.