Link Prediction Algorithms Via Simulated Weight Strategy In Complex Networks

With the advancement of science and technology,network science focusing on complex networks has developed rapidly.As one of the popular research directions in network science,the main task of link prediction is to discover unknown links or incorrect links in the network.The research on link prediction not only helps to understand the network structure and the evolution mechanism,but also has a wide range of social application values.Most of the existing indexes are based on the similarity of nodes to measure the possibility of generating links.In contrast,fewer indexes are proposed from the perspective of the importance of links.Therefore,taking the unweighted and undirected network as the research object,this thesis proposes a simulated weight strategy for the links based on the idea of weighting the nodes or links in the network dynamics.Combined with this strategy,new prediction algorithms are proposed from three perspectives: link topology,node similarity and network representation learning.Firstly,combined the strategy with the topology of the edge clustering coefficient,a simulated weight index based on the link clustering coefficient,ERA,is proposed.The results of experiments on 10 real networks show that the ERA index has a good predictive effect,and the results also verify that the edge clustering coefficient can directly reflect the local structure of the network.Secondly,in order to further improve the prediction accuracy,this paper consider using the similarity of nodes to describe the importance of links.At the same time,combined with the resource allocation,this paper uses information compensation strategy to distinguish the contribution of neighbor nodes at different distances,and proposes three prediction indexes(RCN,RAA and RRA).The prediction accuracy under different evaluation indexes is analyzed through the experiments,and the good prediction performance and stability of the new indexes are verified.Finally,in order to reduce the time consumption,based on the network representation learning,this paper uses node space vectors to describe the intimacy between nodes,and proposes a new prediction model.And it combines DeepWalk,LINE and Node2 vec three representation learning algorithms to propose corresponding prediction indexes,which effectively improve the computational efficiency.The experimental results show that these prediction indexes have higher prediction accuracy.In addition,compared with LINE and Node2 vec,DeepWalk-based weight simulation prediction indexes have a better prediction performance and robustness.