Link Prediction Via Modular Structures And Network Representation Learning

Complex networks are widely used in modelling relational data in recent years.Consider an undirected graph with vertices denoting entities and edges representing relationships between entities,the essential goal of link prediction lies in seeking the primal causes of the edges generations.Besides,link prediction plays an indefensible role in investigating the evolution of complex networks or doing analytic characteristics on structured data.Additionally,link prediction has been applied broadly in various scientific fields,such as guiding the direct interactions validation in biological networks,or recommending friendships in human social networks.However,due to the high complexity of network structures,the requirements of dealing with massive volumes networks and handling intricate relations in complex network considerably increase.On the other hand,the uncertainty caused by noises,and the sparsity in network also trigger some difficult issues in link prediction.In general,according to the above mentioned problems,the attentions are paid mainly in preserving informative structures in complex networks.Specifically,with limited,intricate,and sparse network,how to achieve as much structure features as possible is the key concern in the dissertation.The main contributions are as follows:1)As for group evolution,the proposed model combines the node features from inter-cluster and intra-cluster to predict the missing link interactions in real-world networks.Compared with other models,it obtains much better performance,especially in human protein network.Furthermore,the robustness of the proposed model to noise is high enough Even the ratio of missing links occupies to 50%,this model can provide sustainable prediction performance.2)Proposes a fusion model of low rank and sparse representation for link prediction.We take the advantages of both the low rank representation and sparse representation to fit the interactions between node pairs.From the experimental results,the proposed model owns competitive performances in predicting edges.3)Describes the modular structures in network with a confined low rank representation models.By delineating relationships from intra-class and inter-class simultaneously,an approximation for network with modular structures can be obtained.From experimental results,the proposed model can compete with other baseline algorithms in predicting accuracies.4)Puts forward a manifold regularized low rank representation in link prediction model,which can unitedly reconstruct network from both global and local structures.The sparsity of the network itself results in a low rank recovery for observed matrix,the similarity structures of nodes are preserved by locality invariance.Sufficient validations have been experimented,and the results show that the proposed model is more competitive compared with almost all other existing methods.In particular,in the weighted real networks performs more superior prediction precisions to other available algorithms.5)High level features from deep neural models are proved to be effective for discriminative similarity extraction from complex networks.In order to capture more accurate features network process,refined non-linear features integrated with self-representation is proposed to capture more structures in network and an improved link prediction accuracy can be obtained.Meanwhile,amounts of experiments are implemented to show to verify the feasibility of the proposed model.In conclusion,this dissertation focuses on managing un-weighted complex network,and considers how to represent the Modular structures in network,how to describe relationships simultaneously from Inter-class and Intra-class in a combinatory representation,how to comprehensively approximate network structures from global and local perspectives in a flexible way,and how to combine Deep Learning and Traditional machine learning models to constitute a feature extraction scheme for link prediction.Meanwhile,amounts of experiments are implemented to verify the feasibility of all the methods proposed in this dissertation and the advantages compared with the existing methods are illustrated in detail.