Applying GAN-based Methods To Data Completion

High quality data is the premise and guarantee that data can be fully mined and reliable conclusions can be drawn.The lack of data will have a negative impact on the subsequent data analysis and mining.Only by reasonably completing the incomplete data can people obtain the theories,methodologies and technologies that support follow-up research,management and decision-making from the data.Therefore,it is of great significance to complete the incomplete data for the subsequent data analysis and mining.To increase the accuracy of data completion,this paper proposes two novel methods based on GANs: dual discriminators generative adversarial imputation nets pre trained by latent factor model（LFM-D2GAIN）and attention reconstruction mechanism-based end-to-end generative adversarial nets（ARM-E2GAN）.These two methods aim at static data completion and dynamic data completion,respectively.The contributions of this paper are as follows:Firstly,in the scenario of static data completion,LFM applies pre-training steps to the GAIN.At the same time,the structure of the original network is improved by adding another discriminator with different functions.Secondly,in the scenario of dynamic data completion,the ARM-E2 GAN adds the prefilling step to the E2 GAN,uses the short-time real sequence to train the discriminator,integrates the attention mechanism into the decoding process of the generator,and redistributes the attention scores of each hidden state in the encoder,so as to reconstruct the attention.Finally,the above two methods and five baseline methods were used to carry out control experiments in three missing rates,two data types and eight different data sets.Under the RMSE index,the results obtained by the LFM-D2 GAIN are reduced by 21.19% on average compared with other methods.Among them,the results obtained in static data completion are reduced by32.95% on average compared with other methods;Compared with other methods,the results of ARM-E2 GAN increase by 12% on average,and the results obtained in dynamic data completion decrease by 20.36% on average.The two methods proposed in this paper can achieve the expected goal in static and dynamic data completion respectively,which is helpful to improve the effect of data completion.In addition,in order to achieve better completion effect,it is necessary to select the appropriate data completion method according to the type of the data.