Design For Time Series Imputation Scheme Based On Generative Model

With the development of data science and artificial intelligence technology,people's demand for data has increased,and a large amount of data has been applied in different scenarios.Time Series Data,as an important data type,is very common in real life.It is of great practical significance to mine the hidden information in the time series data and analyze the time series data.However,due to the instability of the data collection equipment or human reasons,the collected data is not complete data,and there are missing values,which affects the follow-up research on time series data.The existing missing value processing methods include direct deletion method,traditional missing value imputation method and missing value imputation method based on generative model.However,these methods are either only suitable for sample sets with a small missing rate,or they do not take into account the time sequence information in the time series data,so the effect of the missing value imputation problem in the time series data is not satisfactory.In order to improve the shortcomings of the existing imputation methods in the processing of missing values of time series data,this paper proposes two high-performance time series data missing value imputation algorithms based on two excellent generative models.The specific work is as follows:Designed a time series data missing value imputation scheme based on the Generative Adversarial Network(GAN)model.The Multi-head Self Attention mechanism(MSA)is used in the Generator,so that it can learn the timing information in the time series data well,avoiding the traditional Recurrent Neural Network(RNN)dealing with the problem of long sequences easily causing forgetting.The construction of the generator uses the Auto-Encoder(AE)structure,so that the initial input of the generator is no longer a random vector,but a more excellent feature vector obtained by the encoder compression.There are imputation methods that have improved the imputation accuracy.Designed a time series data missing value imputation scheme based on the Variational AutoEncoder(VAE)model.Both the encoder and the decoder reference the newest Prob Sparse Self Attention mechanism(PSA),which has the same advantage as the traditional Self Attention mechanism(SA)in dealing with long sequence problems.Its computational complexity is smaller than that of traditional SA.This method not only improves the imputation accuracy compared with the existing imputation methods,but also improves the time efficiency under the premise of not much difference in accuracy compared with the GAN-based method.In order to verify the effectiveness of the two methods proposed in this thesis,two real time series data set is selected,and the existing method and the two methods in this article are used to impute the missing values.The results show that the method proposed in this paper is significantly better than the existing methods in terms of imputation performance.