Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis

Han Liu; Xiaolin Yang; Di Wu; Xiaobin Yuan; Ji Zhang; Rafal Kustra

Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis

Han Liu^*, Xiaolin Yang, Di Wu, Xiaobin Yuan, Ji Zhang, Rafal Kustra

^*Corresponding author for this work

Computer Science

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

Recently, semi-supervised classification has drawn more attention and many practical semi-supervised learning methods have been proposed. However, current literature ignores an important fact-How to estimate the exact labeled sample size given a lot of unlabeled samples, which is important because of the rareness and expensiveness of labeled examples and is also crucial for us to explore the relative value of labeled and unlabeled samples given a specific model. Based on the assumption of a latent gaussian-distribution to the domain, we describe a reasonable labeled sample size estimation method for semi-supervised linear discriminant analysis (Transductive LDA). A detailed mathematical derivation and a computationally tractable approach are given out. Our technique extends naturally to handle two difficult problems: learning from gaussian distributions with different covariances, and learning for multiple classes.

Original language	English (US)
Title of host publication	Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 and Proceedings of the International Conference on Machine Learning; Models, Technologies and Applications, MLMTA'04)
Editors	H.R. Arabnia, M. Youngsong
Pages	1007-1012
Number of pages	6
State	Published - 2004
Event	Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 - Las Vegas, NV, United States Duration: Jun 21 2004 → Jun 24 2004

Publication series

Name	Proceedings of the International Conference on Artificial Intelligence, IC-AI'04
Volume	2

Other

Other	Proceedings of the International Conference on Artificial Intelligence, IC-AI'04
Country/Territory	United States
City	Las Vegas, NV
Period	6/21/04 → 6/24/04

Keywords

Bayes risk
Sample size estimation
Semi-supervised classification
Transductive LDA
Unlabeled data

ASJC Scopus subject areas

Artificial Intelligence

Cite this

Liu, H., Yang, X., Wu, D., Yuan, X., Zhang, J., & Kustra, R. (2004). Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis. In H. R. Arabnia, & M. Youngsong (Eds.), Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 and Proceedings of the International Conference on Machine Learning; Models, Technologies and Applications, MLMTA'04) (pp. 1007-1012). (Proceedings of the International Conference on Artificial Intelligence, IC-AI'04; Vol. 2).

Liu, Han ; Yang, Xiaolin ; Wu, Di et al. / Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis. Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 and Proceedings of the International Conference on Machine Learning; Models, Technologies and Applications, MLMTA'04). editor / H.R. Arabnia ; M. Youngsong. 2004. pp. 1007-1012 (Proceedings of the International Conference on Artificial Intelligence, IC-AI'04).

@inproceedings{3c8937b2cbc24278b2fdd76006746ff4,

title = "Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis",

abstract = "Recently, semi-supervised classification has drawn more attention and many practical semi-supervised learning methods have been proposed. However, current literature ignores an important fact-How to estimate the exact labeled sample size given a lot of unlabeled samples, which is important because of the rareness and expensiveness of labeled examples and is also crucial for us to explore the relative value of labeled and unlabeled samples given a specific model. Based on the assumption of a latent gaussian-distribution to the domain, we describe a reasonable labeled sample size estimation method for semi-supervised linear discriminant analysis (Transductive LDA). A detailed mathematical derivation and a computationally tractable approach are given out. Our technique extends naturally to handle two difficult problems: learning from gaussian distributions with different covariances, and learning for multiple classes.",

keywords = "Bayes risk, Sample size estimation, Semi-supervised classification, Transductive LDA, Unlabeled data",

author = "Han Liu and Xiaolin Yang and Di Wu and Xiaobin Yuan and Ji Zhang and Rafal Kustra",

year = "2004",

language = "English (US)",

isbn = "1932415335",

series = "Proceedings of the International Conference on Artificial Intelligence, IC-AI'04",

pages = "1007--1012",

editor = "H.R. Arabnia and M. Youngsong",

booktitle = "Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 and Proceedings of the International Conference on Machine Learning; Models, Technologies and Applications, MLMTA'04)",

note = "Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 ; Conference date: 21-06-2004 Through 24-06-2004",

}

Liu, H, Yang, X, Wu, D, Yuan, X, Zhang, J & Kustra, R 2004, Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis. in HR Arabnia & M Youngsong (eds), Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 and Proceedings of the International Conference on Machine Learning; Models, Technologies and Applications, MLMTA'04). Proceedings of the International Conference on Artificial Intelligence, IC-AI'04, vol. 2, pp. 1007-1012, Proceedings of the International Conference on Artificial Intelligence, IC-AI'04, Las Vegas, NV, United States, 6/21/04.

Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis. / Liu, Han; Yang, Xiaolin; Wu, Di et al.
Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 and Proceedings of the International Conference on Machine Learning; Models, Technologies and Applications, MLMTA'04). ed. / H.R. Arabnia; M. Youngsong. 2004. p. 1007-1012 (Proceedings of the International Conference on Artificial Intelligence, IC-AI'04; Vol. 2).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis

AU - Liu, Han

AU - Yang, Xiaolin

AU - Wu, Di

AU - Yuan, Xiaobin

AU - Zhang, Ji

AU - Kustra, Rafal

PY - 2004

Y1 - 2004

N2 - Recently, semi-supervised classification has drawn more attention and many practical semi-supervised learning methods have been proposed. However, current literature ignores an important fact-How to estimate the exact labeled sample size given a lot of unlabeled samples, which is important because of the rareness and expensiveness of labeled examples and is also crucial for us to explore the relative value of labeled and unlabeled samples given a specific model. Based on the assumption of a latent gaussian-distribution to the domain, we describe a reasonable labeled sample size estimation method for semi-supervised linear discriminant analysis (Transductive LDA). A detailed mathematical derivation and a computationally tractable approach are given out. Our technique extends naturally to handle two difficult problems: learning from gaussian distributions with different covariances, and learning for multiple classes.

AB - Recently, semi-supervised classification has drawn more attention and many practical semi-supervised learning methods have been proposed. However, current literature ignores an important fact-How to estimate the exact labeled sample size given a lot of unlabeled samples, which is important because of the rareness and expensiveness of labeled examples and is also crucial for us to explore the relative value of labeled and unlabeled samples given a specific model. Based on the assumption of a latent gaussian-distribution to the domain, we describe a reasonable labeled sample size estimation method for semi-supervised linear discriminant analysis (Transductive LDA). A detailed mathematical derivation and a computationally tractable approach are given out. Our technique extends naturally to handle two difficult problems: learning from gaussian distributions with different covariances, and learning for multiple classes.

KW - Bayes risk

KW - Sample size estimation

KW - Semi-supervised classification

KW - Transductive LDA

KW - Unlabeled data

UR - http://www.scopus.com/inward/record.url?scp=12744281771&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=12744281771&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:12744281771

SN - 1932415335

SN - 9781932415339

T3 - Proceedings of the International Conference on Artificial Intelligence, IC-AI'04

SP - 1007

EP - 1012

BT - Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 and Proceedings of the International Conference on Machine Learning; Models, Technologies and Applications, MLMTA'04)

A2 - Arabnia, H.R.

A2 - Youngsong, M.

T2 - Proceedings of the International Conference on Artificial Intelligence, IC-AI'04

Y2 - 21 June 2004 through 24 June 2004

ER -

Liu H, Yang X, Wu D, Yuan X, Zhang J, Kustra R. Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis. In Arabnia HR, Youngsong M, editors, Proceedings of the International Conference on Artificial Intelligence, IC-AI'04 and Proceedings of the International Conference on Machine Learning; Models, Technologies and Applications, MLMTA'04). 2004. p. 1007-1012. (Proceedings of the International Conference on Artificial Intelligence, IC-AI'04).

Statistical issues with labeled sample size analysis for semi-supervised linear discriminant analysis

Abstract

Publication series

Other

Keywords

ASJC Scopus subject areas

Other files and links

Fingerprint

Cite this