Link Prediction Algorithm Research Based On Neighborhood Smoothing

With the increasing application of network data or graph data in complex systems,network link prediction has always been an important research content in the fields of network science and data mining.This paper uses neighborhood smoothing to provide a new framework for network link prediction,proposes a neighborhood smoothing algorithm,and uses evaluation methods to analyze the prediction results,which effectively solves the problems of high complexity and strong structural hypothesis in network prediction,and has good prediction effect.Link prediction has made many achievements in similarity-based methods,and has been effectively and widely used in various networks such as social networks,biological networks,communication networks and other networks.However,these often only consider the local information of the network,such as the classic similarity index based on common neighbors,and how to further use the network topology information remains to be studied.The framework based on likelihood analysis has better link prediction accuracy,but it also has its shortcomings,and the computational complexity is too high to be applied to large-scale networks.The framework based on neighborhood smoothing proposed in this paper effectively considers the global of the network,innovatively obtains the prediction edge probability matrix on the estimation of neighborhood smoothing,and uses the adaptive smoothing method to predict the network link.The main research contents of this paper include:First,in the algorithm of network link prediction,the previous prediction methods based on similarity and maximum likelihood estimation,as well as the evaluation indicators of network prediction,are studied and introduced,which is based on the framework of neighborhood smoothing.Secondly,in the technical research of neighborhood smoothing,starting from the estimation of the edge probability matrix,the expectation and similarity score matrix of the network adjacency matrix are used to directly carry out the probability estimation,which effectively unifies the existing methods for modeling.Then,in the neighborhood selection of neighborhood smoothing,the distance between nodes is defined based on the Graphon function,and the node set with the smaller distance value is adaptively selected as the neighborhood set according to the nature of the network itself,so that the smoothed estimation is more accurate.This distance definition is then extended to the high-order distance between nodes for neighborhood selection.Neighborhood selection based on similarity forms the neighborhood set of neighboring nodes with large similarity,and the node sets obtained by different similarity algorithms are different and different in the process of smoothing calculation.In addition,the prediction methods on some special networks,this paper takes symbolic networks and temporal networks as examples,and uses the framework of neighborhood smoothing to carry out theoretical research.Finally,various network link prediction methods based on neighborhood smoothing are compared and analyzed,and good results are shown on different real networks,and the prediction accuracy of neighborhood smoothing based on similarity is improved in most results than that of probability estimation using only the similarity score matrix.Based on the network link prediction method of neighborhood smoothing,this paper mainly carries out the prediction analysis on general static networks,and the innovation lies in the network adaptive neighborhood selection,and nodes change their link probability with the local network structure change,rather than only changing the link or attribute indicators of nodes.Experiments show that this neighbor smoothing link prediction method can effectively improve the accuracy of network prediction.In the future,the evolution of directed networks,weighted networks and dynamic networks using effective neighborhood selection to explore the law of network evolution needs to be further studied.