Robust learning of the modified Huber loss

This paper investigates pairwise robust learning based on the modified Huber loss function. Robustness is achieved at the cost of sacrificing learning rates. The modified Huber loss, while maintaining robustness, effectively enhances convergence rates compared to the exponential squared loss. Pairwise learning is a machine learning approach that focuses on learning from pairwise examples. Within a reproducing kernel Hilbert space, we establish a comparison theorem and derive two concentration inequalities, thereby determining the convergence rates of the model, which optimally reaches [Formula: see text].