Error Correcting Output Coding (ECOC), an information theoretic concept, seems an attractive idea for improving the performance of automatic classifiers, particularly for problems that involve large number of classes. Converting a complex multi-class problem to a few binary problems allows the use of less complex learning machines, that are then combined by assigning the class according to closest distance to a code word defined by the ECOC matrix. We look at the conditions necessary for reduction of error in the ECOC framework and introduce a new version of ECOC called circular ECOC which is less sensitive to code word selection. To demonstrate the error reduction process and compare the two algorithms, we design an artificial benchmark on which we are able to control the rate of noise and visualize the decision boundary to investigate behavior in different parts of input space. Experimental results on a few popular real data bases are also presented to reinforce our conclusions.
[1]
Leo Breiman,et al.
Bagging Predictors
,
1996,
Machine Learning.
[2]
Reza Ghaderi,et al.
AdaBoost and neural networks
,
1999,
ESANN.
[3]
Thomas G. Dietterich,et al.
Solving Multiclass Learning Problems via Error-Correcting Output Codes
,
1994,
J. Artif. Intell. Res..
[4]
Yoav Freund,et al.
Boosting a weak learning algorithm by majority
,
1990,
COLT '90.
[5]
Robert Tibshirani,et al.
Classification by Pairwise Coupling
,
1997,
NIPS.
[6]
Thomas G. Dietterich,et al.
Error-Correcting Output Coding Corrects Bias and Variance
,
1995,
ICML.
[7]
Charles L. Wilson,et al.
Binary decision clustering for neural-network-based optical character recognition
,
1994,
Pattern Recognit..