Study On Link Prediction Based On Machine Learning Methods

Link prediction is a hot topic in the field of complex networks and an important method for studying the relationships or possible connections between nodes.It plays a very important role in the analysis of many complex networks such as protein interaction networks,food chain networks,etc.Traditional link prediction algorithms mainly utilize the similarity features between nodes,but a single similarity indicator lacks comprehensive consideration of the relationships between node pairs.Therefore,this article delves into how to better utilize this information to improve the accuracy and efficiency of predictions.First,this paper analyzes link prediction based on a single similarity metric.Eight similarity indicators were selected to carry out link prediction experiments on six public network datasets of Karate,Jazz,Celegans,USair,Dolphins and Facebook,and the AUC value of the prediction based on a single similarity index was obtained,which was convenient for the comparative analysis between algorithms later.Secondly,the link prediction effect based on supervised learning algorithm is studied by transforming the link prediction problem into a binary classification problem.In order to make full use of the similarity characteristics between nodes,this paper takes eight similarity indicators between nodes as feature vectors between nodes,and uses Logistic regression algorithm and XGBoost ensemble algorithm to experiment on the above six datasets and compare with a single similarity index algorithm.The results show that the XGBoost algorithm achieves high AUC values on the four datasets of Karate,Jazz,Dolphins and Facebook,and the logistic regression algorithm achieves higher AUC values on the Celegans dataset.Finally,in order to obtain the feature vectors between nodes more conveniently and efficiently,this article uses the Node2 vec algorithm to extract the neighborhood information of nodes to obtain the feature vectors,and uses graph convolutional neural networks and graph attention neural networks for lotus path prediction,respectively.The Node2 vec based graph convolutional neural network(N-GCN)link prediction algorithm and the Node2 vec based graph attention neural network(N-GAT)link prediction algorithm are obtained.By analyzing and comparing the AUC values of different link prediction methods through example analysis,it was found that the AUC values of link prediction methods based on N-GCN and NGAT were better than those of supervised machine learning algorithms based on a single similarity index.The N-GCN link prediction algorithm achieved the optimal AUC values on Karate,Jazz,Celegans,and USair networks,while the N-GAT prediction algorithm achieved the optimal AUC values on Dolphins and Facebook networks.