Research On Link Attribute Prediction Technology Of Complex Network

In real life,all kinds of complex systems that people come into contact with can be abstracted into complex networks.Meanwhile,complex networks involve many different fields,so the research on complex networks also attracts researchers from different disciplines.Researching and analyzing these diverse networks will help us understand the real-world social,economic,technological and biological systems.There are many problems related to complex network,this paper mainly explores the related technologies of link prediction.Because link prediction can reveal potential hidden attribute relationships in the network,it has important research significance and value.Most of the existing link prediction algorithms are based on the network topology,because the network structure information has reliability and credibility.However,the existing algorithms are limited by the accuracy and generality,so it is very important to build a more general and high-precision link prediction algorithm.This paper studies the unweighted and weight networks,The main research results are as follows:NCE model is a representation learning algorithm combined with matrix factorization which is suitable for link prediction in undirected and unweighted netowrks.It solves the problem of graph node embedding from the perspective of measuring the global probability transfer between any tow nodes in the network.The Global Probability Transfer Matrix(GPTM)calculation method between two nodes is defined here,It is a combination two methods of random walk and matrix decomposition to define the association index between nodes and apply it to the learning process of network embedding.Here we will look at the unweighted link prediction and supervised binary classification problem.The node's representation vector is combined with a logistic regression model to predict the probability of a link between any nodes.The accuracy and reliability of the NCE model were verified by AUC and Precision indicators on 8 real network datasets.The robustness of the NCE model to sparse networks with different degrees is analyzed.The NCE model is analyzed by combining four mainstream graph representation learning methods with the generalization ability with node embedding dimensions and the time complexity of training and learning.We propose a general link prediction architecture NEW for undirected weighted networks.The characteristic properties of each node are defined by combining the neighborhood information of each node.Here,we regard the link weight prediction problem as a supervised regression problem.In the framework of supervised learning,we use the defined characteristic attributes to solve the link weight prediction problem.For this reason,we propose a polynomial function to fit the link strength.Through RMSE and PCC indicators on six real network datasets,it is verified that NEW model has obvious advantages over other baseline methods.At the same time,the stability and reliability of NEW model on Sparse network are analyzed.In order to show that NEW model can deal with large-scale networks,we theoretically prove that the time complexity of the model is approximately linear,which shows that the model can be easily extended to the problems caused by the linear or exponential growth of the network scale with time evolution.In addition,the role of free hyper-parameter introduced in the NEW model is further revealed.We have made some comprehensive analysis and found parameter optimization can actually make the predicted weights have a closer center and distribution with the actual weights statistically.