Sequential Data Imputation with Evolving Generative Adversarial Networks

Data imputation is an important and widely researched problem in data analysis with applications to a variety of problem domains. Existing literature explores different methods to estimate missing values in a dataset. In this paper, we analyze the task of imputing multiple incomplete numerical datasets that arrive sequentially, each with an overlapping yet distinguishable set of features.We propose a novel procedure based on the Generative Adversarial Network (GAN) architecture, called EvoGAN, which uses transfer learning to build a unified model to perform imputation for all datasets in the sequence. Our methods exploit overlap between datasets and utilize the knowledge gathered from previous rounds. We validate our model on several sequences of datasets and find compelling results for compression rate and acceleration which outperform the state-of-the-art imputation models that are based on an input of a single dataset.