Feature selection under regularized orthogonal least square regression with optimal scaling

Due to lack of scale change in orthogonal least square regression (OLSR), the scaling term is introduced to OLSR to build up a novel orthogonal least square regression with optimal scaling (OLSR-OS) problem in this paper. In addition, the proposed OLSR-OS problem is proven to be numerically better than the OLSR problem. In order to select relevant features under the proposed OLSR-OS problem, ℓ2, 1-norm regularization is further introduced, such that row-sparse projection is achieved. Accordingly, a novel parameterized expansion balanced feature selection (PEB-FS) method is derived based on an extension balanced counterpart. Moreover, not only the convergence of the proposed PEB-FS method is provided but the optimal scaling can be automatically achieved as well. Consequently, the effectiveness and the superiority of the proposed PEB-FS method are verified both theoretically and experimentally.