Learning a powerful SVM using piece-wise linear loss functions

TL;DR

This paper introduces a flexible SVM framework that learns optimal piece-wise linear loss functions tailored to the training data, unifying and improving upon existing SVM variants.

Contribution

The paper proposes a novel k-piece-wise linear loss SVM model that adapts to data and generalizes several existing SVM models, with extensive experimental validation.

Findings

k-PL-SVM outperforms traditional SVMs in experiments

The model adapts loss functions to data characteristics

Existing SVMs are special cases of the proposed model

Abstract

In this paper, we have considered general k-piece-wise linear convex loss functions in SVM model for measuring the empirical risk. The resulting k-Piece-wise Linear loss Support Vector Machine (k-PL-SVM) model is an adaptive SVM model which can learn a suitable piece-wise linear loss function according to nature of the given training set. The k-PL-SVM models are general SVM models and existing popular SVM models, like C-SVM, LS-SVM and Pin-SVM models, are their particular cases. We have performed the extensive numerical experiments with k-PL-SVM models for k = 2 and 3 and shown that they are improvement over existing SVM models.