基于图卷积神经网络的分布式半监督自动标注方法

Distributed and semi-supervised automatic annotation method based on graph convolutional neural network

为解决数据中台建设中数据共享融合难与综合应用难等突出问题,基于图卷积神经网络,面向分布式数据中台架构开发一种分布式的半监督标注算法。通过分析数据缺失、数据量大以及通信带宽限制等问题,首先使用子空间学习提出基于新的正则化项的半监督学习策略。在此基础上,进一步提出基于二级中台的分布式可解释注意力图卷积神经网络融合方法,以解释边的重要性,从而在每一层图神经网络基于私有模型与每个平台的公共数据进行融合,提升模型精准性。在公共数据集上的广泛实验结果表明,该方法可以有效提高标签预测的效果,对大数据背景下的分布式融合、应用具有一定指导意义。

In order to solve the prominent problems such as difficult data sharing and fusion and comprehensive application in the construction of data center,this paper develops a distributed semi-supervised annotation algorithm for distributed data center architecture based on graph convolutional neural network.By analyzing the problems of missing data,large data amount and communication bandwidth limitation,subspace learning is first used to propose a semi-supervised learning strategy based on new regularization terms.On this basis,a distributed explanatory attention graph convolutional neural network fusion method based on the secondary middle platform is further proposed to explain the importance of edges,so that the graph neural network integrates the private model and the public data of each platform to improve the accuracy of the model.Extensive experimental results on public datasets show that it can effectively improve the effect of label prediction and has certain guiding significance for distributed fusion and application in the context of big data.