Doubly Sparse Factor Models For Unifying Feature Transformation And Feature Selection

A number of unsupervised learning methods for high-dimensional data are largely divided into two groups based on their procedures, i.e., (1) feature selection, which discards irrelevant dimensions of the data, and (2) feature transformation, which constructs new variables by transforming and mixing over all dimensions. We propose a method that both selects and transforms features in a common Bayesian inference procedure. Our method imposes a doubly automatic relevance determination (ARD) prior on the factor loading matrix. We propose a variational Bayesian inference for our model and demonstrate the performance of our method on both synthetic and real data.