L1 Regression with Lewis Weights Subsampling

Abstract
We consider the problem of finding an approximate solution to regression while only observing a small number of labels. Given an unlabeled data matrix , we must choose a small set of rows to observe the labels of, then output an estimate whose error on the original problem is within a factor of optimal. We show that sampling from according to its Lewis weights and outputting the empirical minimizer succeeds with probability for . This is analogous to the performance of sampling according to leverage scores for regression, but with exponentially better dependence on . We also give a corresponding lower bound of .
View on arXivComments on this paper