Utility Analysis Of Factors Affecting Network Links

The behaviors of complex networks,such as the evolution direction,the scales emergence,and the egroup coordination behaviors,depend not only on the behavior of individuals,but also on the interactions between individuals.The main research contents of link prediction is the interaction between two individuals.It shows the formation and evolution of the network from the micro level.In addition,link prediction also plays an important role in the fields of recommendation systems,scientific cooperation research and biological science.In short,the research of link prediction has a great significance both in theory and applications.So far,the link prediction algorithms are verious,but none of them are fully applicable to all kinds of networks.Therefore,the existing challenge is to propose a strategy which can help users to select the appropriate link prediction algorithm for a specific network.This paper discusses the important factors affecting the link generation: the endpoint activity,the similarity between nodes,the connection distance and the connection strength,then finds the main factor to help users to select or reconstruct an appropriate link prediction algorithm:(1)We analyse the impact of endpoint activity on links.Traditional algorithms focus on connection relationship,but ignore the influence of the endpoint activity.We found that the more active the endpoint is,the more likely it is to generate a link through experimental analysis.What's more,the degree centrality and pagerank have a more significant impact on links.Therefore,a class of link prediction methods based on endpoint activity is proposed whose accuracy is much higher than the original ones.(2)Whether attribute similarity leads to links? A lot of experiments have been carried out in the paper,and it is found that the similarity between nodes leads to links in assortative networks;the difference leads to links in disassortative networks;the similarity or difference has negligible impact on links in a neutral network.This evolutionary mechanism leads that attribute distribution is more concentrated and attribute heterogeneity is lower in assortative and disassortative networks,while the attribute distribution of the neutral network is more dispersed and heterogeneous.(3)We analyze the impact of connection relationship on link prediction.The connection relationship includes distance and the connection strength between nodes.The experiments show that the shorter the distance is,the greater the connection strength is,the more likely links are generated.Subsequently,the traditional indexes of link prediction are improved by adding the connection relationship between nodes.This paper not only analyzes the factors affecting link prediction,but also provides an effective strategy for selecting link prediction algorithms.Users can use simple elements in link prediction,such as endpoint activity,node similarity,connnection distance,connection strength or other possible factors,and find the main factors affecting link generation,then choose or reconstruct new link prediction methods.In summary,this paper further enriches the link prediction system in the views of basic element analysis and link prediction algorithm selection strategy.The work of this paper provides a new idea for link prediction problems,and is valuable in the fields of recommendation system and social computing.