Local scaling heuristic-based regularization for pattern classification

• It proposes a local scaling heuristic-based regularization (LSHR) for binary classification.
• This LSHR can reflect the intra-class compactness and inter-class separability of outputs.
• The H-LSHR and LS-LSHR classifiers with the hinge and least squares loss functions are presented based on LSHR.

In this paper, a novel regularization method called the local scaling heuristic-based regularization (LSHR) is proposed for binary classification. The idea in LSHR is to integrate the underlying knowledge inside the training points, including the intra-class and inter-class local information in training points. By combining the local scaling heuristic strategy, this LSHR uses two matrices defined on the intra-class and inter-class graphs of points to reflect the intra-class compactness and inter-class separability of outputs. Based on the LSHR method, two classifiers with the hinge and least squares loss functions, H-LSHR and LS-LSHR, are presented for binary classification. The experimental results on several artificial, UCI benchmark datasets and USPS digit datasets indicate the effectiveness of the proposed method.