Link Prediction Algorithm Based On Supervised Learning

Link prediction connects complex systems and information science.It deals with the most basic problem in information science-the reduction and prediction of missing information.In general,supervised algorithms tend to achieve better performance in the field of link prediction than unsupervised algorithms.However,due to the particularity of complex networks,supervised algorithms in the field of link prediction will face two major problems: classification imbalance and feature extraction.For the two problems,the dissertation makes the following contributions in the field of link prediction from the aspects of unbalanced classification and supervised feature extraction:1.A Netboost-LP(Netboost Link Prediction)algorithm is proposed for the imbalanced classification in link prediction.Based on the application of traditional Adaboost algorithm in link prediction,a bias adjustment function is added for complex networks.This adjustment changes the rule of updation of weights and data connection mode of the original algorithm.In the training process,according to the existing learning level of the model,the algorithm adaptively changes the data distribution to achieve the purpose of balancing the data classification.And the effect of the bias adjustment function on the convergence of the loss function is proved.Finally,through the prediction of four real network links,the conclusion that Netboost-LP algorithm has better performance in large networks is verified.2.A supervised feature extraction method for imbalanced problem is proposed.According to the similarity between complex network and text processing research,a link prediction algorithm based on word vector and particle swarm optimization is proposed,which ispired by the deep learning feature extraction algorithm in text processing and particle swarm optimization in optimization problem.This method transforms the abstract network topology relationship into a mathematical expression pattern of display.Then,under the supervision,the particle swarm optimization algorithm is used to filter the extracted features,and to determine the resampling parameters for solving the imbalanced classification problem.The computational complexity of different link prediction algorithms is analyzed.Finally,the algorithm of this paper is compared with three kinds of link prediction algorithms based on similarity,deep learning and unbalanced data in different time series networks.The results show that the proposed link prediction algorithm has higher prediction accuracy.The algorithm is more stable and more universal.3.The two supervised link prediction algorithms proposed in this dissertation are integrated into one system and applied to the network invulnerability problem under fuzzy information.Two parameters of network ambiguity and prediction ratio are designed.When the information is incomplete,the parameters are used to assist in formulating the corresponding deliberate attack strategy based on link prediction.The experimental results show that the attack strategy formulated after link prediction is better than the direct attack under fuzzy information when there is less missing information.Even in some specific networks,it will be better than direct attacks under full information.This shows that the supervised link prediction algorithm in this dissertation reveals the intrinsic evolution mechanism of the network from a certain perspective.