Data-Driven Pinball-Loss Selection for Elastic-Net SVMs

Pinball-loss support vector machines (SVMs) are known for their robustness, but their performance often depends on manually pre-setting an asymmetry parameter. This research introduces a novel data-driven elastic-net SVM that automatically learns simplex-constrained weights over various candidate pinball losses. This effectively creates a pinball loss with a data-dependent parameter, all while maintaining a single, unified classifier. For high-dimensional datasets, the paper proposes a column-partitioned variable-splitting solver. This solver ensures convergence and, under specific conditions, produces identical iterates and solutions to centralized training, even across different column partitions. Experiments confirm its predictive behavior, numerical equivalence, and scalability across multiple processes, offering a more adaptive and efficient approach to robust classification.