Research On Link Prediction Algorithm Based On Network Representation Learning

With the rapid development of network science,the link prediction has become an important subject in complex network.According to the given network structure,node properties and so on,link prediction predicts the probability that a link may exist between unconnected nodes in the network.Network representation learning can map network nodes to a low-dimensional vector space,capture the topological properties of the network and reduce the time and space complexity.This paper carries out the link prediction algorithms based on network representation learning aims at the homogeneous network datasets and heterogeneous network datasets.The main contents are as follows:(1)Most traditional link prediction algorithms are designed based on the adjacency matrix of networks.These have higher calculation complexity,can't retain higher-order construct of network and have a risk of falling into the local minimum value due to initializing randomly.Therefore,this paper proposed a link prediction algorithm based on generative adversarial hierarchical networks representation learning(GAHNRL).Firstly,according to the first-order proximity and second-order proximity of the network,the method recursively performs edge collapsing and vertex merging on the network to form multi-layered sub-networks with a smaller layer-by-layer scale.Secondly,Node2 vec algorithm is used to pre-process the smallest sub-network,and the pre-processed result is input into generative adversarial networks to learn low-dimensional vector representation of the vertices in the smallest sub-network.Thirdly,the learned low-dimensional vector representation of the vertices in the smallest sub-network is input into the generative adversarial networks of the previous sub-network to learn the vector representation of the vertices in the previous sub-network.According to this method,learning process is recursively layer-by-layer back-up until the original network is processed,and a low-dimensional vector representation of all vertices is obtained.Extensive experiments performed on real network datasets in multiple different fields demonstrate that the proposed metric can outperform the other six state-of-the-art metrics.The proposed algorithm can retain higher-order construct of network and avoid the partial minimum problem.(2)Since the traditional link prediction algorithms are interested only a relationship type of network,these can't adapt to the heterogeneous networks.So this paper proposed a heterogeneous networks representation learning link prediction algorithm based on scene driven(Scene2vec).Firstly,the heterogeneous network is processed as a whole to obtain base vector representation.According to scene identifiers,the scenes of heterogeneous networks is divided to form a scene network.Secondly,the scene network is input into the autoencoder to learn the vector representation of each scene.The scene driven vector representation is obtained by performing weighted accumulation on all scene vector representations of each node.Finally,the low-dimensional vector representation is obtained by linearly combining the basic vector representation and the scene driven vector representation.Extensive experiments performed on heterogeneous networks demonstrate that the Scene2 vec can outperform state-of-the-art metrics.The Scene2 vec can retain the network structure of the original network,node similarity and scene similarity.And it can capture the scene driven first-order proximity and second-order proximity among different type nodes on heterogeneous networks.