Understanding and improving error-correcting output coding Public Deposited

http://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/wh246z090

Descriptions

Attribute NameValues
Creator
Abstract or Summary
  • Error-correcting output coding (ECOC) is a method for converting a k-classsupervised learning problem into a large number L of two-class supervised learningproblems and then combining the results of these L evaluations. Previous researchhas shown that ECOC can dramatically improve the classi cation accuracy of supervisedlearning algorithms that learn to classify data points into one of k 2 classes.An investigation of why the ECOC technique works
  • Error-correcting output coding (ECOC) is a method for converting a k-classsupervised learning problem into a large number L of two-class supervised learningproblems and then combining the results of these L evaluations. Previous researchhas shown that ECOC can dramatically improve the classi cation accuracy of supervisedlearning algorithms that learn to classify data points into one of k 2 classes.An investigation of why the ECOC technique works, particularly when employedwith decision tree learning algorithms, is presented.It is shown that the ECOC method is a compact form of voting amongmultiple hypotheses. The success of the voting depends on that the errors committedby each of the L learned binary functions are substantially uncorrelated.By employing the statistical notions of bias and variance, the generalizationerrors of ECOC are decomposed into bias and variance errors. Like any votingmethod, ECOC reduces variance errors. However, unlike homogeneous voting, whichsimply combines multiple runs of the same learning algorithm, ECOC can alsoreduce bias errors. It is shown that the bias errors in the individual functions are uncorrelated and that this results from non-local behavior of the learning algorithmin splitting the feature space.ECOC is also extended to provide class probability information. The problemof computing these class probabilities can be formulated as an over-constrained systemof linear equations. Least squares methods are applied to solve these equations.Accuracy of the posterior probabilities is demonstrated with overlapping classes anda simple reject option task.
Resource Type
Date Available
Date Copyright
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Academic Affiliation
Non-Academic Affiliation
Keyword
Rights Statement
Language
Replaces
Additional Information
  • description.provenance : Submitted by Philip Vue (vuep@onid.orst.edu) on 2009-05-07T15:56:26Z No. of bitstreams: 1 Eun_Bae_Kong.pdf: 779245 bytes, checksum: 25c381cefebbf6df9cc36ecafdc609a1 (MD5)
  • description.provenance : Approved for entry into archive by Linda Kathman(linda.kathman@oregonstate.edu) on 2009-05-12T16:38:42Z (GMT) No. of bitstreams: 1 Eun_Bae_Kong.pdf: 779245 bytes, checksum: 25c381cefebbf6df9cc36ecafdc609a1 (MD5)
  • description.provenance : Approved for entry into archive by Linda Kathman(linda.kathman@oregonstate.edu) on 2009-05-12T16:41:22Z (GMT) No. of bitstreams: 1 Eun_Bae_Kong.pdf: 779245 bytes, checksum: 25c381cefebbf6df9cc36ecafdc609a1 (MD5)
  • description.provenance : Made available in DSpace on 2009-05-12T16:41:22Z (GMT). No. of bitstreams: 1 Eun_Bae_Kong.pdf: 779245 bytes, checksum: 25c381cefebbf6df9cc36ecafdc609a1 (MD5)

Relationships

Parents:

This work has no parents.

Last modified

Downloadable Content

Download PDF

Items