Feature Selection Using Autoencoders

Feature selection plays a vital role in improving the generalization accuracy in many classification tasks where datasets are high-dimensional. In feature selection, a minimal subset of relevant as well as non-redundant features is selected. Autoencoders are used to represent the datasets from original feature space to a reduced and more informative feature space. In this paper, we propose a novel approach for feature selection by traversing back the autoencoders through more probable links. Experiments on five publicly available large datasets show that our approach gives significant gains in accuracy over most of the state-of-the-art feature selection methods.