Neighborhood-aware autoencoder for missing value imputation
Helena Aidos, Pedro Tomás · 2020
Missing values are a fundamental issue in many applications by constraining the application of different learning methods or by impairing the attained results. Many solutions have been proposed by relying on statistical or machine learning techniques. However, in most cases, the results are not yet satisfactory. Hence, motivated by the advent of deep learning, different solutions have also been proposed, such as by adopting autoencoders and adversarial training. However, in most of these solutions, the results are impaired by the network structure and training strategy, constraining the accuracy of missing value imputation. In this paper, we revisit autoencoder networks and show that through a careful selection of network structure and optimization strategy we outperform other deep learning solutions. We further study the impact of a previously proposed technique, stochastic corruption of inputs, to show that when the network is well designed and trained, it actually impairs the results.