1: \begin{abstract}

2: This paper deals with sparse feature selection and grouping for

3: classification and regression. The classification or

4: regression problems under consideration consists in minimizing 

5: a convex empirical risk

6: function subject to an $\ell^1$ constraint, a pairwise

7: $\ell^\infty$ constraint, or a pairwise

8: $\ell^1$ constraint. Existing work, such as the Lasso

9: formulation, has focused mainly on Lagrangian penalty

10: approximations, which often require ad hoc or computationally

11: expensive procedures to determine the penalization parameter. 

12: We depart from this approach and

13: address the constrained problem directly via a splitting method.

14: The structure of the method is that of the classical 

15: gradient-projection algorithm, which alternates a 

16: gradient step on the

17: objective and a projection step onto the lower level set modeling

18: the constraint. The novelty of our approach is that the 

19: projection step is implemented via an outer approximation scheme 

20: in which the constraint set is approximated by

21: a sequence of simple convex sets consisting of the intersection of

22: two half-spaces. Convergence of the iterates generated by the

23: algorithm is established for a general smooth convex minimization

24: problem with inequality constraints. Experiments on both synthetic and

25: biological data show that our method outperforms penalty methods. 

26: \end{abstract}

27: