Abstract: Logistic regression is a promising method that has been used widely in many applications of data mining, machine learning, computer vision. In this paper, we study on the l₀ sparse logistic regression problem. This problem has been proposed as a promising method for feature selection in classification problems, and it is in general NP-hard. In order to solve this problem, we develop a splitting augmented Lagrange method with embedding BB (Barzilai and Borwein) method (SALM-BB). At each iteration, SALM-BB solves an unconstrained convex subproblem and a quadratic programming with l₀ constraint. We use BB method to solve the unconstrained convex subproblem. For the quadratic programming subproblem, we obtain its exact optimal solution directly. Furthermore, we prove the convergence of SALM-BB, under certain assumptions. Finally, we implement SALM-BB on the l₀ sparse logistic regression problem based on the simulation data and the data of University of California, Irvine, Machine Learning Repository. Numerical results show that SALM-BB outperforms the well-known first-order solver SLEP in terms of lower average logistic loss and lower misclassification error rate.

Key words: sparse logistic regression, splitting augmented Lagrange method, feature selection