Scientific Computing Seminar

Date:

Friday, September 3, 2004

Time:

12:00pm-1:00pm

Location:

50D-3416

Seminar Speaker:

Jie Zhou
Department of Computer Science
Northern Illinois University

Title:

Designing Multi-class Classifiers Using Error Correcting Output Codes

Abstract:

Multi-class classifiers are very useful in many areas of pattern recognition and data mining such as bioinformatics, image classification, etc. Multiple binary classifiers such as the popular Support Vector Machines can be combined to achieve multiclass classification. Commonly used approaches include the one-vs-others scheme and the one-vs-one (pairwise coupling) scheme. Often, a quadratic number of base learners are required to obtain good results. We present a novel approach to design multi-class classifiers based on error correcting output codes (ECOC) which associate each class with a codeword. This approach allows a unified framework of traditional schemes. We study one kind of ECOC classifier that decodes using minimum Hamming distance and designs codewords based on inter-cluster distance. Experiments are conducted on both synthetic data and real world applications including UCI repository problems and CENPARMI handwritten numerals. Results show that the proposed scheme can achieve competitive accuracy compared with traditional schemes, and the number of base learners is typically much less than that required by the pairwise scheme.

Sponsor of Seminar:

Chris Ding

Scientific Computing

Contact Esmond G. Ng EGNg@lbl.gov