基于改进RNN的多变量时间序列缺失数据填充算法

Multivariate missing data imputing algorithm based on modified RNN

随着大数据时代的来临,多变量时间序列的应用价值得到了越来越多的关注。然而,缺失数据的存在严重影响了对多变量时间序列的进一步开发利用。针对这个问题,提出了基于改进递归神经网络的多变量缺失数据填充算法,该算法通过衰减机制可以获得更多有用的隐藏信息,从而更好地完成对多变量缺失数据的填充。首先,对多变量数据进行预处理,得到网络的输入向量;其次,在长短时记忆(Long-Short-Term Memory,LSTM)单元的基础上引入衰减机制,提出了两种改进的缺失数据填充模型。改进后的模型能够更多更好地获取长时间间隔的隐藏信息,并对输入进行相应的衰减处理。为检验算法的性能,在上海空气质量数据集以及多传感器数据融合活动识别系统(Activity Recognition system based on Multisensor data fusion,AReM)数据集上进行了仿真实验。结果表明,相比于其他算法,所提算法能够更好地实现多变量时间序列的缺失数据填充。

More and more attention has been paid to the application value of the multivariate time series with the coming of the era of big data.The existence of missing data,however,seriously affected the further development and utilization of multivariate time series.In the view of this problem,a multivariate missing data imputing algorithm is proposed based on modified recurrent neural network,which can get more useful hidden information through the damping mechanism,so as to complete multivariate missing data imputing better.First of all,the imput vector of the network is obtained by preprocessing multivariate data.Secondly,two modified multivariate missing data imputing models are proposed via introducing the damping mechanism on the basis of the Long-Short-Term Memory(LSTM).The improved models can access the hidden information of the long time interval more and cope with the input vector by the damping mechanism.Experiments are carried out on the Shanghai air quality dataset and the Activity Recognition system based on Multisensor data fusion(AReM)dataset to verify the performance of the algorithm.The experimental results show that the proposed algorithm can realize the multivariate missing data imputing better compared with other algorithms.