The Research On Link Prediction Related Issues Based On Machine Learning

Many complex systems in real world can be represented by networks,where the nodes represent entities and links capture the relationship between them.As one of the basic but fundamental problems in complex network,link prediction is still an open issue and many related problems are constantly emerging.Link prediction has widespread practical values in different fields and has always been a hotspot of interdisciplinary research.Recently,machine learning,especially deep leaning,has made great breakthrough in Euclidean structure data.Meantime,using deep models to deal with non-Euclidean structure data such as networks is just unfolding.In this thesis,link prediction and its related issues are studied in depth and three improved algorithms based on machine learning have been proposed to solve the following issues:(1)Link prediction based on hierarchical attention mechanism: Traditional similarity indices are difficult to extract effective structural information in sparse networks.Inspired by graph attention network,we propose an end-to-end link prediction model based on hierarchical attention mechanism.This model,consisting of two levels of attention mechanism including nodes and links,can adaptively learn the effective structural information.Experiments in many real networks,especially in Internet and Power network,show that the proposed method significantly outperforms the baseline methods.Besides,we validate the robustness of the proposed method by limiting the size of training set.(2)Link weight prediction based on feature fusion of multi-network: Different from link prediction which aims to predict the existence of links in networks,we sometimes need to predict the link weight in weighted networks due to the missing information.Therefore,we propose a link weight prediction algorithm based on feature fusion of multi-network.Specifically,we first transform the original network into line graph and then define the link centrality in original network by node centrality in line graph.Then we put the similarity indices and link centrality together to train the traditional leaning algorithms.The experimental results in many real networks show that the proposed method has better prediction effect than the baseline algorithms.Further,we find that the line graph has complementary effect on link weight prediction(3)Link concealment based on evolutionary algorithm: In the respective of privacy and security,we propose the counterpart problem of link prediction,namely link concealment.We regard link prediction as a tool of attackers to acquire sensitive links,the defenders need to take some protective measures to avoid leakage of privacy links.Two link perturbation mechanisms based on evolutionary algorithms are proposed in this thesis.Moreover,a method based on incremental updating is proposed to accelerate the calculation of fitness.Empirical experiments in several real networks show that the proposed method has the best defensive effect compared with several baseline algorithms.In addition,the proposed method has good defensive transferability against different link prediction attacks.