Abstract
The best-known decomposition schemes of multiclass learning problems are one per class coding (OPC) and error-correcting output coding (ECOC). Both methods perform a prior decomposition, that is, before training of the classifier takes place. The impact of output codes on the inferred decision rules can be experienced only after learning. Therefore, we present a novel algorithm for the code design of multiclass learning problems. This algorithm applies a maximum-likelihood objective function in conjunction with the expectation-maximization (EM) algorithm. Minimizing the augmented objective function yields the optimal decomposition of the multiclass learning problem in two-class problems. Experimental results show the potential gain of the optimized output codes over OPC or ECOC methods.
Original language | English |
---|---|
Pages (from-to) | 1065-1102 |
Number of pages | 38 |
Journal | Neural Computation |
Volume | 13 |
Issue number | 5 |
DOIs | |
State | Published - May 2001 |