Unsupervised feature selection based on non-parametric mutual information

We present a novel filter approach to unsupervised feature selection based on the mutual information estimation between features. Our feature selection approach does not impose a parametric model on the data and no clustering structure is estimated. Instead, to measure the statistical dependence between features, we employ a mutual information criterion, which is computed by using a non-parametric method, and remove uncorrelated features. Numerical experiments on synthetic and real world tasks show that the performance of our algorithm is comparable to previously suggested state-of-the-art methods.