Unsupervised Feature Selection using Encoder-Decoder Networks

摘要

Feature selection is one of the most important techniques for dimension reduction in a wide range of tasks. This paper presents a simple yet efficient method for unsupervised feature selection. To learn an encoder-decoder network on a huge number of training samples, the initial weights of such network must be regularly updated until the convergence reached. We interestingly investigated that the weights of an input neuron to the latent space neurons are highly correlated with the importance of them. We tailor an efficient measure as the importance of each feature value relies on the changing of connected weights to it. The experimental results confirm that the performance of proposed method is comparable, even better than other state-of-the-art methods.