Unsupervised Feature Recommendation using Representation Learning

Today's world extensively depends on analytics of high dimensional sensor time-series, and, extracting informative representation. Sensor time-series across various applications such as healthcare and human wellness, machine maintenance etc., are generally unlabelled, and, getting the annotations is costly and time-consuming. Here, we propose an unsupervised feature selection method exploiting representation learning with a choice of best clustering and recommended distance measure. Proposed method reduces the feature space, to a compressed latent representation, known as Auto-encoded Compact Sequence of features, by retaining the most informative parts. It further selects a set of discriminative features, by computing the sim-ilarity / dissimilarity among the features in latent space using the recommended best distance measure. We have experimented using diverse time-series from UCR Time Series Classification archive, and observed, proposed method consistently outperforms state-of-the-art feature selection approaches.