Adaptive thresholding and reweighting to improve domain transfer learning for unbalanced data with applications to EEG imbalance

Kyung Min Su, Kay A. Robbins, W. David Hairston

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Domain adaptation methods can be highly sensitive to class balance, particularly the usually unknown balance of the unlabeled test set. In this work, we analyze the effect of imbalance on a well-known algorithm, ARTL (Adaptation Regularization Transfer Learning) and propose four approaches for mitigating the adverse effects of imbalance. These include (1) balancing the training set for pseudo-label calculation, (2) applying adaptive thresholding to pseudo-label calculation, (3) using class reweighting in the optimization objective, and (4) applying adaptive thresholding to the output objective. We tested these methods with the UCI newsgroup dataset and on three types of imbalanced EEG (electroencephalogram) classification problems. We observed significant improvements, particularly for cases of extreme imbalance, which are not well addressed by standard classification techniques.

Original languageEnglish (US)
Title of host publicationProceedings - 2016 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages320-325
Number of pages6
ISBN (Electronic)9781509061662
DOIs
StatePublished - Jan 31 2017
Event15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016 - Anaheim, United States

Other

Other15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016
CountryUnited States
CityAnaheim
Period12/18/1612/20/16

Fingerprint

Electroencephalography
Labels

Keywords

  • Adaptive threshold
  • Imbalance
  • Transfer Learning

ASJC Scopus subject areas

  • Artificial Intelligence
  • Computer Networks and Communications
  • Computer Science Applications

Cite this

Su, K. M., Robbins, K. A., & Hairston, W. D. (2017). Adaptive thresholding and reweighting to improve domain transfer learning for unbalanced data with applications to EEG imbalance. In Proceedings - 2016 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016 (pp. 320-325). [7838163] Institute of Electrical and Electronics Engineers Inc.. DOI: 10.1109/ICMLA.2016.34

Adaptive thresholding and reweighting to improve domain transfer learning for unbalanced data with applications to EEG imbalance. / Su, Kyung Min; Robbins, Kay A.; Hairston, W. David.

Proceedings - 2016 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016. Institute of Electrical and Electronics Engineers Inc., 2017. p. 320-325 7838163.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Su, KM, Robbins, KA & Hairston, WD 2017, Adaptive thresholding and reweighting to improve domain transfer learning for unbalanced data with applications to EEG imbalance. in Proceedings - 2016 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016., 7838163, Institute of Electrical and Electronics Engineers Inc., pp. 320-325, 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016, Anaheim, United States, 18-20 December. DOI: 10.1109/ICMLA.2016.34
Su KM, Robbins KA, Hairston WD. Adaptive thresholding and reweighting to improve domain transfer learning for unbalanced data with applications to EEG imbalance. In Proceedings - 2016 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016. Institute of Electrical and Electronics Engineers Inc.2017. p. 320-325. 7838163. Available from, DOI: 10.1109/ICMLA.2016.34

Su, Kyung Min; Robbins, Kay A.; Hairston, W. David / Adaptive thresholding and reweighting to improve domain transfer learning for unbalanced data with applications to EEG imbalance.

Proceedings - 2016 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016. Institute of Electrical and Electronics Engineers Inc., 2017. p. 320-325 7838163.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

@inbook{d1c5aae1f08249dfb8c76f07222aea5b,
title = "Adaptive thresholding and reweighting to improve domain transfer learning for unbalanced data with applications to EEG imbalance",
abstract = "Domain adaptation methods can be highly sensitive to class balance, particularly the usually unknown balance of the unlabeled test set. In this work, we analyze the effect of imbalance on a well-known algorithm, ARTL (Adaptation Regularization Transfer Learning) and propose four approaches for mitigating the adverse effects of imbalance. These include (1) balancing the training set for pseudo-label calculation, (2) applying adaptive thresholding to pseudo-label calculation, (3) using class reweighting in the optimization objective, and (4) applying adaptive thresholding to the output objective. We tested these methods with the UCI newsgroup dataset and on three types of imbalanced EEG (electroencephalogram) classification problems. We observed significant improvements, particularly for cases of extreme imbalance, which are not well addressed by standard classification techniques.",
keywords = "Adaptive threshold, Imbalance, Transfer Learning",
author = "Su, {Kyung Min} and Robbins, {Kay A.} and Hairston, {W. David}",
year = "2017",
month = "1",
doi = "10.1109/ICMLA.2016.34",
pages = "320--325",
booktitle = "Proceedings - 2016 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016",
publisher = "Institute of Electrical and Electronics Engineers Inc.",
address = "United States",

}

TY - CHAP

T1 - Adaptive thresholding and reweighting to improve domain transfer learning for unbalanced data with applications to EEG imbalance

AU - Su,Kyung Min

AU - Robbins,Kay A.

AU - Hairston,W. David

PY - 2017/1/31

Y1 - 2017/1/31

N2 - Domain adaptation methods can be highly sensitive to class balance, particularly the usually unknown balance of the unlabeled test set. In this work, we analyze the effect of imbalance on a well-known algorithm, ARTL (Adaptation Regularization Transfer Learning) and propose four approaches for mitigating the adverse effects of imbalance. These include (1) balancing the training set for pseudo-label calculation, (2) applying adaptive thresholding to pseudo-label calculation, (3) using class reweighting in the optimization objective, and (4) applying adaptive thresholding to the output objective. We tested these methods with the UCI newsgroup dataset and on three types of imbalanced EEG (electroencephalogram) classification problems. We observed significant improvements, particularly for cases of extreme imbalance, which are not well addressed by standard classification techniques.

AB - Domain adaptation methods can be highly sensitive to class balance, particularly the usually unknown balance of the unlabeled test set. In this work, we analyze the effect of imbalance on a well-known algorithm, ARTL (Adaptation Regularization Transfer Learning) and propose four approaches for mitigating the adverse effects of imbalance. These include (1) balancing the training set for pseudo-label calculation, (2) applying adaptive thresholding to pseudo-label calculation, (3) using class reweighting in the optimization objective, and (4) applying adaptive thresholding to the output objective. We tested these methods with the UCI newsgroup dataset and on three types of imbalanced EEG (electroencephalogram) classification problems. We observed significant improvements, particularly for cases of extreme imbalance, which are not well addressed by standard classification techniques.

KW - Adaptive threshold

KW - Imbalance

KW - Transfer Learning

UR - http://www.scopus.com/inward/record.url?scp=85015408611&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85015408611&partnerID=8YFLogxK

U2 - 10.1109/ICMLA.2016.34

DO - 10.1109/ICMLA.2016.34

M3 - Conference contribution

SP - 320

EP - 325

BT - Proceedings - 2016 15th IEEE International Conference on Machine Learning and Applications, ICMLA 2016

PB - Institute of Electrical and Electronics Engineers Inc.

ER -