Study Of Network Representation Learning Based On Neural Networks

Network representation learning aims to analyze large-scale network problems,and it has attracted wide attention from many researchers in recent years.Although there have been many effective models for network representation learning,there remain some problems unresolved.For example,many methods do not take the heterogeneous networks into consideration and not use the label information of the networks sufficiently.With the development and growth of the neural network models,these models have shown great performance in the fields of image,audio,video,text processing and etc.How to apply it to network representation learning needs further exploration.To address the above problems,this paper carried out in-depth study and research,and completed the following work:This thesis innovatively apples the Generative Adversarial Networks to the representation learning of heterogeneous networks,and proposes an effective preprocessing method for the decomposition of heterogeneous networks according to different edge types.Through the adversarial learning of different edge types in the network,the unsupervised network representation learning model of this paper can effectively capture the overall structure information of heterogeneous networks and the connection relationship between different types of nodes.For the heterogeneous networks with label information,this paper constructs a semisupervised network representation learning model.By transforming the label information into the connection relationship between nodes,the label information is fully utilized.It can effective extract and represent different characteristic types of nodes and edges in heterogeneous networks.At the same time,a semi-supervised clustering algorithm is proposed,which constructs the label information as constraints to guide the clustering.This work performs experiments on three public dataset to compare this work with state-of-the-art methods.And the experimental results show that the adversarial based method and the semi-supervised method can effective extract and represent network characteristics,and obtain better results than other methods.