Abstract—The application of robust loss function is an important approach to classify data sets that contaminated by noisy data points, in particular by outliers. In this paper we present an extension of smoothed 0-1 loss function to the multiclass case. In multiclass case, Fisher consistency of smoothed 0-1 loss function is satisfied. A classification algorithm is developed for multiclass classification problems. The performance of Hinge loss function and smoothed 0-1 loss function based classification algorithms are compared on several data sets with different levels of noise. Experiments show that smoothed 0-1 loss function demonstrates improved performance for data classification on more noisy datasets with noisy features or labels.
Index Terms—Optimization, classification, loss function, robust.
Lei Zhao is with the Office of Research, University of the Sunshine Coast, Maroochydore QLD 4558 Australia (e-mail: lzhao@usc.edu.au).
Cite:Lei Zhao, "A Robust Loss Function for Multiclass Classification," International Journal of Machine Learning and Computing vol.3, no. 6, pp. 462-467, 2013.