A robust self-weighted SELO regression model

Abstract

Linear regression model is a useful tool in machine learning and has been applied in diverse fields including compressed sensing, computer vision and matrix analysis, among many others. For linear regression model, variable selection and parameter estimation are the most fundamental and important tasks. The seamless-$$L_{0}$$ penalty estimator (SELO), which can finish variable selection and parameter estimation simultaneously, is attractive for its good theoretical properties and easy computation. However, the SELO is sensitive to outliers. Besides, seamless-$$L_{0}$$ is non-convex, so most of the existing computing methods may easily converge to (bad) local minima. To solve these problems, a novel robust self-weighted model of SELO (RSWSELO) is proposed for linear regression. The RSWSELO is proved to be consistent in term of parameter estimation and variable selection, like the SELO model, which means the RSWSELO can converge to oracle estimator (asymptotically equivalent to the least squares estimator constrained to the true nonzero coefficients). An adaptive regularizer is introduced to the proposed model, which can assign weights to the selected samples based on the loss of samples during the iteration process. Thus, the weights can be decided by the model automatically and the proposed model is much more robust than the SELO model. Furthermore, the experimental results on simulation studies and UCI datasets demonstrate that the proposed model is effective and outperforms SELO in generalization performance.