CoChat
Sign Up
Cited by 921Imbalanced Data Classification TechniquesMachine Learning and Data ClassificationData Mining Algorithms and Applications

Learning when training data are costly: the effect of class distribution on tree induction

Open access
Weiss, G. M., Provost, F.·2003·Journal of Artificial Intelligence Research
DOI 10.1613/jair.1199
Open this paper in CoChat — free accountOpen PDF

Abstract

For large, real-world inductive learning problems, the number of training examples often must be limited due to the costs associated with procuring, preparing, and storing the training examples and/or the computational costs associated with learning from them. In such circumstances, one question of practical importance is: if only n training examples can be selected, in what proportion should the classes be represented? In this article we help to answer this question by analyzing, for a fixed training-set size, the relationship between the class distribution of the training data and the performance of classification trees induced from these data. We study twenty-six data sets and, for each, determine the best class distribution for learning. The naturally occurring class distribution is shown to generally perform well when classifier performance is evaluated using undifferentiated error rate (0/1 loss). However, when the area under the ROC curve is used to evaluate classifier performance, a balanced distribution is shown to perform well. Since neither of these choices for class distribution always generates the best-performing classifier, we introduce a budget-sensitive progressive sampling algorithm for selecting training examples based on the class associated with each example. An empirical analysis of this algorithm shows that the class distribution of the resulting training set yields classifiers with good (nearly-optimal) classification performance.

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1613/jair.1199

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • The optimal distribution for training decision-tree classifiers often involves oversampling the minority class.

    “Moreover, the optimal distribution has an oversampling of the minority class (there are also cases when the best classifier is obtained with an oversampling up to 90% of the minority class samples).”

    From Fame for sale: Efficient detection of fake Twitter followers (2015) · cites this paper as [34]

    Section 6.3 (page unavailable) reports that, for AUC, 11 of the 12 data sets whose optimal range excludes the natural distribution have more minority-class examples in that range. This supports the claim that increasing the minority-class share can be optimal. The cited 90% example is not established as the best classifier overall: the paper says it performs better at high true-positive rates in one example, while a balanced distribution has the highest AUC for that data set. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • The metric used to evaluate the performance of different classifiers changes the optimal distribution of the classes for the training set.

    “The conclusions of their investigation have shown that the metric used to evaluate the performance of the different classifiers changes the optimal distribution of the classes for the training set.”

    From Fame for sale: Efficient detection of fake Twitter followers (2015) · cites this paper as [34]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Cites

Flagged first, then by citations

Research works printed but not linked: 4

  1. Good, I. J. (1965). The Estimation of Probabilities. Cambridge, MA: M.I.T. Press.
  2. Japkowicz, N., Holte, R. C., Ling, C. X., & Matwin S. (Eds.) (2000). In Papers from the AAAI Workshop on Learning from Imbalanced Data Sets. Tech, rep. WS-00-05, Menlo Park, CA: AAAI Press.
  3. Provost, F., Fawcett, T., & Kohavi, R. (1998). The case against accuracy estimation for compar- ing classifiers. In Proceedings of the Fifteenth International Conference on Machine Learning. San Francisco, CA: Morgan Kaufmann.
  4. Turney P. (2000). Types of cost in inductive learning. In Workshop on Cost-Sensitive Learning at the Seventeenth International Conference on Machine Learning, 15-21, Stanford, CA.

Cited by

Sorted by citations

Related

Authors
Weiss, G. M., & Provost, F.
Affiliations
AT&T (United States) (US)
ORCID
Venue
Journal of Artificial Intelligence Research
Publisher
AI Access Foundation
Published
2003-10-01
Volume
19
Pages
315-354
ISSN
1076-9757 (ISSN-L)
DOI
10.1613/jair.1199
OpenAlex
W2114968414
Cited by
921
References
38
Citations by year
2026: 15 · 2025: 25 · 2024: 21 · 2023: 30 · 2022: 36 · 2021: 45 · 2020: 47 · 2019: 39 · 2018: 62 · 2017: 50 · 2016: 42 · 2015: 56 · 2014: 49 · 2013: 60 · 2012: 38
FWCI
28.30
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Journal of Artificial Intelligence Research (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • Faculty Digital Archive (New York University Florence) (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Imbalanced Data Classification Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
  • Physical Sciences › Computer Science › Information Systems › Data Mining Algorithms and Applications
Keywords
class distribution, training data selection, tree induction, classification trees, cost-sensitive learning, ROC curve, 0-1 loss, class imbalance, inductive learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review