Hidden AS link prediction based on random forest feature selection and GWO-XGBoost model

Abstract

Internet AS-level topology measurement is crucial for improving network stability and security. The presence of hidden AS links poses a challenge for accurately measuring the AS-level topology. Link prediction serves as a primary technical approach for discovering hidden AS links. However, the effectiveness of existing methods is susceptible to features and model hyperparameters, necessitating improvements in prediction performance. In this paper, a hidden AS link prediction method based on random forest feature selection and a GWO-XGBoost model is proposed. First, BGP data is preprocessed to eliminate erroneous information from AS paths, and suitable AS triplets for training the prediction model are constructed. Then, the traffic volume and ratio at the first and last nodes of these triplets are analyzed to extract four new features. These are combined with features extracted by typical methods to form an initial prediction feature set. Additionally, the random forest algorithm is used to select initial features, remove redundant features, and construct an optimal feature subset. Finally, the initial prediction model XGBoost is trained using the optimal feature subset, while the Grey Wolf Optimizer (GWO) algorithm is employed to search for optimal hyperparameters, thus constructing a fusion model GWO-XGBoost that achieves hidden AS link prediction. Extensive experiments are conducted on the AS-level topology with 81,998 nodes and 401,925 links collected from RouteViews and RIPE RIS projects. The results show that the proposed method has significant advantages over the typical prediction methods TopoScope and LOC-TopoScope. The prediction accuracy increases by 5.30% and 3.96%, respectively, and the number of discovered hidden AS links increases by 23.76% and 6.08%, respectively.