贝叶斯推理模型Index-GMVAE在随机数据缺失填补中的应用  

Application of Bayesian Inference Model Index-GMVAE in Filling Random Missing Values of Data

作　　者：马超 蔡猛[2] 李建勋[1] MA Chao;CAI Meng;LI Jian-xun(Shanghai Jiaotong University Department of Automation, Shanghai 200240;Luoyang Institute of Electro-optical Equipment, AVIC, Luoyang 471009, China)

机构地区：[1]上海交通大学自动化系,上海200240 [2]中国航空工业集团洛阳电光设备研究所,河南洛阳471009

出　　处：《指挥控制与仿真》2022年第1期32-37,共6页Command Control & Simulation

基　　金：国家自然科学基金(61673265);国家重点研究开发项目(2020Yfc1512203)。

摘　　要：变分自编码器(VAE)推理随机缺失数据还原是近年来一种新兴的数据填补方法,但传统的VAE算法存在后验分布单一,变分推断证据下界收敛慢等问题。因此,提出了一种采用高斯混合后验分布和缺失数据位置掩码推理缺失数据还原的新的贝叶斯推理网络框架——Index-GMVAE。在Mnist手写数字数据集和Adult数据集上分别做了实验验证,仿真结果表明,提出的算法能明显提高证据下界的收敛速度,相比于VAE算法,将缺失数据推理还原的准确率提高了10%左右,且在图像数据集和异构二分类数据集上均有效提高了模型的概率隐变量层的推理性能,体现了对不同结构数据的普适性。Variational autoencoder(VAE)infers random missing data reduction is an emerging data filling method in recent years,but the traditional VAE algorithm has the problems of single posterior distribution and slow convergence of the evidence lower bound of variational inference.A new Bayesian inference network framework,Index-GMVAE,based on Gaussian mixture posterior distribution and missing data location mask to infer missing data restoration is proposed in this paper.The experiments are performed on the Mnist handwritten digit dataset and the Adult dataset.The simulation results show that the proposed algorithm can significantly improve the convergence speed of the evidence lower bound.Compared with the VAE algorithm,the accuracy of inferring missing data restoration is increased by about 10%.And it can effectively improve the inference performance of the probability hidden layer of model on both the image dataset and the heterogeneous binary classification dataset,which reflects the universality of data with different structures.

关 键 词：变分自编码器 高斯混合模型 缺失位置掩码 证据下界 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]