arXiv · math/0405356
Complexities of convex combinations and bounding the generalization error in classification
Abstract
We introduce and study several measures of complexity of functions from the convex hull of a given base class. These complexity measures take into account the sparsity of the weights of a convex combination as well as certain clustering properties of the base functions involved in it. We prove new upper confidence bounds on the generalization error of ensemble (voting) classification algorithms that utilize the new complexity measures along with the empirical distributions of classification margins, providing a better explanation of generalization performance of large margin classification methods.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Vladimir Koltchinskii, Dmitry Panchenko. 2005-08-25. Complexities of convex combinations and bounding the generalization error in classification. https://doi.org/10.1214/009053605000000228. https://arxiv.science/abs/math/0405356
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.