We give algorithms for regression for a wide class of M-Estimator loss functions. These generalize l_p-regression to fitness measures used in practice such as the Huber measure, which enjoys the robustness properties of l_1 as well as the smoothness properties of l_2. For such estimators we give the first input sparsity time algorithms. Our techniques are based on the sketch and solve paradigm. The same sketch works for any M-Estimator, so the loss function can be chosen after compressing the data.
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.
No comments:
Post a Comment