On regularization parameter estimation under covariate shift

Wouter M. Kouw; Marco Loog

doi:10.1109/ICPR.2016.7899671

On regularization parameter estimation under covariate shift

Wouter M. Kouw, Marco Loog

Department of Computer Science

1 Citation (Scopus)

Abstract

This paper identifies a problem with the usual procedure for L²-regularization parameter estimation in a domain adaptation setting. In such a setting, there are differences between the distributions generating the training data (source domain) and the test data (target domain). The usual cross-validation procedure requires validation data, which can not be obtained from the unlabeled target data. The problem is that if one decides to use source validation data, the regularization parameter is underestimated. One possible solution is to scale the source validation data through importance weighting, but we show that this correction is not sufficient. We conclude the paper with an empirical analysis of the effect of several importance weight estimators on the estimation of the regularization parameter.

Original language	English
Title of host publication	23rd International Conference on Pattern Recognition, ICPR 2016
Number of pages	6
Publisher	IEEE
Publication date	1 Jan 2016
Pages	426-431
ISBN (Electronic)	978-1-5090-4847-2
DOIs	https://doi.org/10.1109/ICPR.2016.7899671
Publication status	Published - 1 Jan 2016
Event	23rd International Conference on Pattern Recognition - Cancun, Mexico Duration: 4 Dec 2016 → 8 Dec 2016 Conference number: 23

Conference

Conference	23rd International Conference on Pattern Recognition
Number	23
Country/Territory	Mexico
City	Cancun
Period	04/12/2016 → 08/12/2016

Access to Document

10.1109/ICPR.2016.7899671

Cite this

@inproceedings{a1bde851ad9843d1a59e4843071cfd83,

title = "On regularization parameter estimation under covariate shift",

abstract = "This paper identifies a problem with the usual procedure for L2-regularization parameter estimation in a domain adaptation setting. In such a setting, there are differences between the distributions generating the training data (source domain) and the test data (target domain). The usual cross-validation procedure requires validation data, which can not be obtained from the unlabeled target data. The problem is that if one decides to use source validation data, the regularization parameter is underestimated. One possible solution is to scale the source validation data through importance weighting, but we show that this correction is not sufficient. We conclude the paper with an empirical analysis of the effect of several importance weight estimators on the estimation of the regularization parameter.",

author = "Kouw, {Wouter M.} and Marco Loog",

year = "2016",

month = jan,

day = "1",

doi = "10.1109/ICPR.2016.7899671",

language = "English",

pages = "426--431",

booktitle = "23rd International Conference on Pattern Recognition, ICPR 2016",

publisher = "IEEE",

note = "23rd International Conference on Pattern Recognition ; Conference date: 04-12-2016 Through 08-12-2016",

}

TY - GEN

T1 - On regularization parameter estimation under covariate shift

AU - Kouw, Wouter M.

AU - Loog, Marco

N1 - Conference code: 23

PY - 2016/1/1

Y1 - 2016/1/1

N2 - This paper identifies a problem with the usual procedure for L2-regularization parameter estimation in a domain adaptation setting. In such a setting, there are differences between the distributions generating the training data (source domain) and the test data (target domain). The usual cross-validation procedure requires validation data, which can not be obtained from the unlabeled target data. The problem is that if one decides to use source validation data, the regularization parameter is underestimated. One possible solution is to scale the source validation data through importance weighting, but we show that this correction is not sufficient. We conclude the paper with an empirical analysis of the effect of several importance weight estimators on the estimation of the regularization parameter.

AB - This paper identifies a problem with the usual procedure for L2-regularization parameter estimation in a domain adaptation setting. In such a setting, there are differences between the distributions generating the training data (source domain) and the test data (target domain). The usual cross-validation procedure requires validation data, which can not be obtained from the unlabeled target data. The problem is that if one decides to use source validation data, the regularization parameter is underestimated. One possible solution is to scale the source validation data through importance weighting, but we show that this correction is not sufficient. We conclude the paper with an empirical analysis of the effect of several importance weight estimators on the estimation of the regularization parameter.

UR - http://www.scopus.com/inward/record.url?scp=85019162968&partnerID=8YFLogxK

U2 - 10.1109/ICPR.2016.7899671

DO - 10.1109/ICPR.2016.7899671

M3 - Article in proceedings

AN - SCOPUS:85019162968

SP - 426

EP - 431

BT - 23rd International Conference on Pattern Recognition, ICPR 2016

PB - IEEE

T2 - 23rd International Conference on Pattern Recognition

Y2 - 4 December 2016 through 8 December 2016

ER -

On regularization parameter estimation under covariate shift

Abstract

Conference

Access to Document

Other files and links

Fingerprint

Cite this