Research On Mining Vital Nodes Based On Reinforcement Learning In Complex Networks

In the field of complex networks,the identification of vital nodes is a hot research issue.For networks with different structures and different types of dynamic processes evolving on the networks,the recognition results of vital nodes often vary greatly.Vital nodes in network dissemination refer to those nodes that can greatly promote the largescale dissemination of information in the network,or when they are immune,they can effectively inhibit the spread of information in the network.Many traditional algorithms for identifying vital nodes mostly rely on a single structural feature of the network.Therefore,most of these algorithms are only suitable for a certain type of network with specific topology.They have greater limitations in terms of universality.Based on the spreading dynamics of the complex network and the topology of the network,this thesis uses the reinforcement learning method to mine vital nodes in the complex networks,and mainly accomplished the following three works of research:Firstly,this thesis proposes a seed nodes selection algorithm for maximum influence based on reinforcement learning.From the perspective of the combination of machine learning and network science,the algorithm studies the problem of maximizing influence on complex networks based on the independent cascade model.Specifically,this method maps the state of the information dissemination network to the characteristic value of the network,thereby transforming complex networks topology information into regular data and serving as input data for reinforcement learning.In this method,the influence of the network topology on the spreading process is fully considered.A large number of numerical simulation experiments have proved that the algorithm can select the most vital node or node set as the seed node on the network of different structures.Secondly,this thesis proposes an optimal control algorithm for rumor spreading on online social networks based on reinforcement learning.Based on the characteristics of rumor spreading,this algorithm proposes a brand-new representation method for the state of the rumor spreading network.At the same time,the influence of the network topology on the spread of rumors is also fully considered.The algorithm can select the most important set of nodes in the rumor spreading network as immune nodes.The simulation results on a large number of synthetic networks and real networks show that the algorithm can effectively curb the number of rumor spreaders and effectively delay the time when the rumors erupt.This work helps people understand the spreading mechanism of rumors and can assist the government to maintain a healthy public opinion environment.Thirdly,we tested the effectiveness of the network state representation method proposed in this paper through the screening method,and further confirmed the feature combination based on the coarse-grained idea.In order to confirm the degree of influence of different characteristics on the learning effect of the model,we used rumor mongering scale and the delay of the outbreak time as reference indicators,and the screening method to verify one by one.a large number of numerical simulations have been performed on the model training on the synthetic network and the real network.Experiments show that the combination of features we selected can effectively represent the network state and has high accuracy.