New analysis and algorithm for learning with drifting distributions

Mehryar Mohri, Andres Muñoz Medina

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

Abstract

We present a new analysis of the problem of learning with drifting distributions in the batch setting using the notion of discrepancy. We prove learning bounds based on the Rademacher complexity of the hypothesis set and the discrepancy of distributions both for a drifting PAC scenario and a tracking scenario. Our bounds are always tighter and in some cases substantially improve upon previous ones based on the L 1 distance. We also present a generalization of the standard on-line to batch conversion to the drifting scenario in terms of the discrepancy and arbitrary convex combinations of hypotheses. We introduce a new algorithm exploiting these learning guarantees, which we show can be formulated as a simple QP. Finally, we report the results of preliminary experiments demonstrating the benefits of this algorithm.

Original languageEnglish (US)
Title of host publicationAlgorithmic Learning Theory - 23rd International Conference, ALT 2012, Proceedings
Pages124-138
Number of pages15
Volume7568 LNAI
DOIs
StatePublished - 2012
Event23rd International Conference on Algorithmic Learning Theory, ALT 2012 - Lyon, France
Duration: Oct 29 2012Oct 31 2012

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume7568 LNAI
ISSN (Print)03029743
ISSN (Electronic)16113349

Other

Other23rd International Conference on Algorithmic Learning Theory, ALT 2012
CountryFrance
CityLyon
Period10/29/1210/31/12

Fingerprint

Discrepancy
Scenarios
Batch
Convex Combination
Experiments
Arbitrary
Experiment
Learning
Standards
Generalization

Keywords

  • domain adaptation
  • Drifting environment
  • generalization bound

ASJC Scopus subject areas

  • Computer Science(all)
  • Theoretical Computer Science

Cite this

Mohri, M., & Muñoz Medina, A. (2012). New analysis and algorithm for learning with drifting distributions. In Algorithmic Learning Theory - 23rd International Conference, ALT 2012, Proceedings (Vol. 7568 LNAI, pp. 124-138). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 7568 LNAI). https://doi.org/10.1007/978-3-642-34106-9_13

New analysis and algorithm for learning with drifting distributions. / Mohri, Mehryar; Muñoz Medina, Andres.

Algorithmic Learning Theory - 23rd International Conference, ALT 2012, Proceedings. Vol. 7568 LNAI 2012. p. 124-138 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 7568 LNAI).

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

Mohri, M & Muñoz Medina, A 2012, New analysis and algorithm for learning with drifting distributions. in Algorithmic Learning Theory - 23rd International Conference, ALT 2012, Proceedings. vol. 7568 LNAI, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 7568 LNAI, pp. 124-138, 23rd International Conference on Algorithmic Learning Theory, ALT 2012, Lyon, France, 10/29/12. https://doi.org/10.1007/978-3-642-34106-9_13
Mohri M, Muñoz Medina A. New analysis and algorithm for learning with drifting distributions. In Algorithmic Learning Theory - 23rd International Conference, ALT 2012, Proceedings. Vol. 7568 LNAI. 2012. p. 124-138. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/978-3-642-34106-9_13
Mohri, Mehryar ; Muñoz Medina, Andres. / New analysis and algorithm for learning with drifting distributions. Algorithmic Learning Theory - 23rd International Conference, ALT 2012, Proceedings. Vol. 7568 LNAI 2012. pp. 124-138 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
@inproceedings{8d7902be98154e048831658c48c12162,
title = "New analysis and algorithm for learning with drifting distributions",
abstract = "We present a new analysis of the problem of learning with drifting distributions in the batch setting using the notion of discrepancy. We prove learning bounds based on the Rademacher complexity of the hypothesis set and the discrepancy of distributions both for a drifting PAC scenario and a tracking scenario. Our bounds are always tighter and in some cases substantially improve upon previous ones based on the L 1 distance. We also present a generalization of the standard on-line to batch conversion to the drifting scenario in terms of the discrepancy and arbitrary convex combinations of hypotheses. We introduce a new algorithm exploiting these learning guarantees, which we show can be formulated as a simple QP. Finally, we report the results of preliminary experiments demonstrating the benefits of this algorithm.",
keywords = "domain adaptation, Drifting environment, generalization bound",
author = "Mehryar Mohri and {Mu{\~n}oz Medina}, Andres",
year = "2012",
doi = "10.1007/978-3-642-34106-9_13",
language = "English (US)",
isbn = "9783642341052",
volume = "7568 LNAI",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
pages = "124--138",
booktitle = "Algorithmic Learning Theory - 23rd International Conference, ALT 2012, Proceedings",

}

TY - GEN

T1 - New analysis and algorithm for learning with drifting distributions

AU - Mohri, Mehryar

AU - Muñoz Medina, Andres

PY - 2012

Y1 - 2012

N2 - We present a new analysis of the problem of learning with drifting distributions in the batch setting using the notion of discrepancy. We prove learning bounds based on the Rademacher complexity of the hypothesis set and the discrepancy of distributions both for a drifting PAC scenario and a tracking scenario. Our bounds are always tighter and in some cases substantially improve upon previous ones based on the L 1 distance. We also present a generalization of the standard on-line to batch conversion to the drifting scenario in terms of the discrepancy and arbitrary convex combinations of hypotheses. We introduce a new algorithm exploiting these learning guarantees, which we show can be formulated as a simple QP. Finally, we report the results of preliminary experiments demonstrating the benefits of this algorithm.

AB - We present a new analysis of the problem of learning with drifting distributions in the batch setting using the notion of discrepancy. We prove learning bounds based on the Rademacher complexity of the hypothesis set and the discrepancy of distributions both for a drifting PAC scenario and a tracking scenario. Our bounds are always tighter and in some cases substantially improve upon previous ones based on the L 1 distance. We also present a generalization of the standard on-line to batch conversion to the drifting scenario in terms of the discrepancy and arbitrary convex combinations of hypotheses. We introduce a new algorithm exploiting these learning guarantees, which we show can be formulated as a simple QP. Finally, we report the results of preliminary experiments demonstrating the benefits of this algorithm.

KW - domain adaptation

KW - Drifting environment

KW - generalization bound

UR - http://www.scopus.com/inward/record.url?scp=84867880979&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84867880979&partnerID=8YFLogxK

U2 - 10.1007/978-3-642-34106-9_13

DO - 10.1007/978-3-642-34106-9_13

M3 - Conference contribution

AN - SCOPUS:84867880979

SN - 9783642341052

VL - 7568 LNAI

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 124

EP - 138

BT - Algorithmic Learning Theory - 23rd International Conference, ALT 2012, Proceedings

ER -