22
19

L1 Regression with Lewis Weights Subsampling

Abstract

We consider the problem of finding an approximate solution to 1\ell_1 regression while only observing a small number of labels. Given an n×dn \times d unlabeled data matrix XX, we must choose a small set of mnm \ll n rows to observe the labels of, then output an estimate β^\widehat{\beta} whose error on the original problem is within a 1+ε1 + \varepsilon factor of optimal. We show that sampling from XX according to its Lewis weights and outputting the empirical minimizer succeeds with probability 1δ1-\delta for m>O(1ε2dlogdεδ)m > O(\frac{1}{\varepsilon^2} d \log \frac{d}{\varepsilon \delta}). This is analogous to the performance of sampling according to leverage scores for 2\ell_2 regression, but with exponentially better dependence on δ\delta. We also give a corresponding lower bound of Ω(dε2+(d+1ε2)log1δ)\Omega(\frac{d}{\varepsilon^2} + (d + \frac{1}{\varepsilon^2}) \log\frac{1}{\delta}).

View on arXiv
Comments on this paper