Adaptive weighted learning for linear regression problems via Kullback–Leibler divergence

In this paper, we propose adaptive weighted learning for linear regression problems via the Kullback–Leibler (KL) divergence. The alternative optimization method is used to solve the proposed model. Meanwhile, we theoretically demonstrate that the solution of the optimization algorithm converges to a stationary point of the model. In addition, we also fuse global linear regression and class-oriented linear regression and discuss the problem of parameter selection. Experimental results on face and handwritten numerical character databases show that the proposed method is effective for image classification, particularly for the case that the samples in the training and testing set have different characteristics.

► We propose adaptive weighted learning for linear regression problems.
► Global and class-oriented weighted linear regression are proposed and fused.
► The problem of parameter selection in our model is discussed.
► Experiments show that the proposed method is effective for image classification.