Time: 2018-12-12 08：00-10：00

Venue: Room 203 Lecture Hall

abstract：Sparse additive models have been successfully applied to high dimensional data analysis due to their representation flexibility and interpretability. However, existing methods are often formulated with the least squares loss under the mean square error (MSE) criterion, which is sensitive to data with the non-Gaussian noise, e.g., the skewed noise, the heavy-tailed noise, and outliers. To cure this problem, we propose a new sparse method, called sparse modal additive model (SpMAM), by integrating the mode-induced loss, the data dependent hypothesis space, and the weighted \ell_{q,1}-norm regularizer (q≥1) into additive models. In contrast to existing methods that aim to learning the conditional mean, the proposed method approximates the intrinsic mode and is robust to the complex noise. Theoretical properties of SpMAM are characterized including generalization bound and variable selection consistency. Experimental results on simulated and benchmark datasets confirm the effectiveness and robustness of the proposed model.