The SVM Classifier with Quartic Truncated Pinball Loss

In the field of machine learning, support vector machine (SVM) is popular for its powerful performance in classification tasks. However, this method could be adversely affected by data containing noise. One of the main reasons is that many loss functions are too sensitive to sample points far from their classes. In this paper, based on the idea of HTPSVM proposed in (Zhu, Song, et al.), we introduce the SVM classifier with quartic truncated pinball loss (QTPSVM), which can be solved by Bregman modified second accelerated proximal gradient method (BAPGs). Numerical experiments show that QTPSVM is more effective on many real data.