Constructing Multiclass Learners from Binary Learners: A Simple Black-Box Analysis of the Generalization Errors
暂无分享,去创建一个
Multiclass learning is widely solved by reducing to a set of binary problems. By considering base binary classifiers as black boxes, we analyze generalization errors of various constructions, including Max-Win, Decision Directed Acyclic Graphs, Adaptive Directed Acyclic Graphs, and the unifying approach based on coding matrix with Hamming decoding of Allwein, Schapire, and Singer, using only elementary probabilistic tools. Many of these bounds are new, some are much simpler than previously known. This technique also yields a simple proof of the equivalences of the learnability and polynomial-learnability of the multiclass problem and the induced pairwise problems.
[1] Thomas P. Hayes,et al. Error limiting reductions between classification tasks , 2005, ICML.
[2] A. Beygelzimer. Multiclass Classification with Filter Trees , 2007 .
[3] Nello Cristianini,et al. Large Margin DAGs for Multiclass Classification , 1999, NIPS.
[4] Boonserm Kijsirikul,et al. Adaptive Directed Acyclic Graphs for Multiclass Classification , 2002, PRICAI.