Title: Classification in the presence of label noise: Structure-aware error bounds via random projections
Authors: Henry Reeve - University of Birmingham (United Kingdom) [presenting]
Ata Kaban - University of Birmingham (United Kingdom)
Abstract: New error bounds for linear classification in conditions of random label noise are presented. To obtain the bounds we use random projections as an analytic device to gain advantage from benign geometric structure that may be present in the data, while working directly with the 0-1 loss. The bounds are data dependent, highlight the characteristics of the problem that make the problem easier or harder, and remain informative in small sample conditions.