Research On Node Influence Ranking Based On Deep Learning

In recent years,complex networks have received more and more attention and research.On the one hand,with the development of information technology,the amount of data we can control has greatly increased;on the other hand,complex networks are models in many fields of reality,and their related theories are of substantial help to grasp the laws of reality.Therefore,more and more scholars in physics and computer science are exploring related theories and technologies based on complex networks.Important node ranking has important applications in various scenarios today,such as viral marketing,virus control,or the spread of rumors in real-world networks.Existing common important node identification methods mainly focus on exploring and defining the importance of nodes through mathematical expressions and heuristic methods.Although heuristic methods have achieved good performance in practice,they are less adaptable,that is,their methods are often only suitable for some specific scenarios and are not universal.Based on deep learning technology,this thesis studies the important node sorting method of complex networks.The main work is as follows:（1）A set of important node identification methods based on deep reinforcement learning is proposed.In this thesis,a node representation method based on the information of the node itself and its neighbor nodes is proposed,and combined with a deep reinforcement learning network,the selection of the seed node set is designed as an"action-reward"-based neural network model.In this thesis,the selection of seed nodes is used as an action,and the difference between the simulated dynamics propagation results of the seed node set is used as a reward,and multiple complex networks are used for training.This thesis finds that the prediction effect of the model is better than that of the algorithm of finding key players in Networks through Deep Reinforcement learning（FINDER）and the algorithm of Incremental Influence Maximization（IMM）.Similar or even larger propagation range,but the prediction time of the method in this thesis is much smaller than these two methods.When the seed set size k is 10,the prediction speed of the algorithm in this thesis in the Live-Journal network is 3.19 times that of the FINDER method.,which is 181 times that of the IMM method;when k is 50,the prediction speed of the algorithm in this thesis in the Live-Journal network is 2.36 times that of the FINDER method and 22.9 times that of the IMM method.And the prediction speed of the algorithm in this thesis also has huge advantages in other networks.（2）An important node sorting algorithm based on Support Vector Regression（Support Vector Regression,SVR）is proposed.This also includes a network embedding method based on feature engineering that combines node degree and extension coreness,and a data sampling method based on eigenvectors.Through the training of the SVR model,this thesis adds the obtained node influence score to the influence of a certain weight of node neighbors to obtain the final node influence score.Finally,the nodes are sorted based on this influence score.The experimental results show that the correlation between the algorithm in this thesis and the real value is relatively stable in each data set,and the Kendall coefficient of the 8 data sets exceeds 70%,which is better than several benchmark algorithms;the uniqueness of the ranking is also the best.,achieves a completely unique ranking on the Denaunay dataset.The model is used to predict node rankings in a non-training network.The correlation between the ranking and the true value is still the best in small and medium-sized networks.The degree of correlation has decreased in large networks,but it is still at the upstream level in all methods;The degree of uniqueness of ranking has decreased compared with the prediction results under the same network,but it is still at the upstream level of all methods.