C4.5, Class Imbalance, and Cost Sensitivity: Why Under-Sampling beats Over-Sampling

By C. Drummond, Robert Craig Holte. 2003

1,123 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

A systematic study of the class imbalance problem in convolutional neural networks (2017, background): “ajority classes until all classes have the same number of examples. While it might not appear intuitive, there is some evidence that in some situations undersampling can be preferable to oversampling [39]. A signicant disadvantage of this method is that it discards a portion of available data. To ov”
A Survey of Predictive Modeling on Imbalanced Domains (2016, background): “The ﬁrst one removes data from the original dataset, reducing the This may increase the likelihood of overﬁtting, especially for higher over-sampling rates [Chawla et al. 2002; Drummond and Holte 2003].”
Using Random Forest to Learn Imbalanced Data (2004, unknown): “As recent research shows (e.g., Ling & Li (1998),Kubat & Matwin (1997),Drummond & Holte (2003)), for the tree classifier, artificially making class priors equal either by down-sampling the majority class or over-sampling the minority class is usually more effective with respect to a given…”
Mind the gaps: weighting the unknown in large-scale one-class collaborative filtering (2009, background): “The two strategies that are commonly used to solve the problem are sampling to re-balance the data [3, 11] and cost-sensitive learning [8, 19].”
Learning Deep Representation for Imbalanced Classification (2016, background): “Therefore under-sampling is often preferred to over-sampling [11], although potentially valuable information may be removed.”
Effective detection of sophisticated online banking fraud on extremely imbalanced data (2012, methodology): “Because the number of genuine transactions is huge, we use under sampling to reduce the data volume and class distribution imbalance [27].”
A study of the behavior of several methods for balancing machine learning training data (2004, background): “Drummond and Holte [8] report that when using C4.”
Striking the Right Balance With Uncertainty (2019, methodology): “These methods include techniques that speciﬁcally focus on class imbalance learning ( P a nd a [ 61 ] AN e t [ 33 ] D ee p I D 2 [ 46 ] O v e r-s a m p li ng [ 14 ] U nd e r-s a m p li ng [ 14 ] T h r e s ho l d A d j .”
Learning from Imbalanced Data (2009, background): “However, this commonality is only superficial, each method introduces its own set of problematic consequences that can potentially hinder learning [25], [38], [39].”
Imbalanced Deep Learning by Minority Class Incremental Rectification (2018, methodology): “by down-sampling majority classes, over-sampling minority classes, or some combinations [5,6,7].”
SMOTE for high-dimensional class-imbalanced data (2013, methodology): “In practice undersampling is effective in removing the gap between the class-specific predictive accuracies for high-dimensional data [7] and it is often used as a reasonable baseline for algorithmic comparison [35].”
Minority report in fraud detection: classification of skewed data (2004, methodology): “Recent work on skewed data sets was evaluated using better performance metrics such as Area Under Curve (AUC) [9, 10], cost curves [15], and Receiver Operating Characteristic (ROC) analysis [28].”
Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect (2020, unknown): “It can be either achieved by re-sampling [25, 26, 15, 16, 27] or re-weighting [13, 14, 12, 17].”
Learned lessons in credit card fraud detection from a practitioner perspective (2014, background): “Undersampling (Drummond & Holte, 2003) consists in down-sizing the majority class by removing observations at random until the dataset is balanced.”
Adaptive Machine Learning for Credit Card Fraud Detection (2015, methodology): “Within the family of sampling methods, we have functions for random undersampling (ubUnder) and oversampling (ubOver) [91].”
Calibrating Probability with Undersampling for Unbalanced Classification (2015, background): “Undersampling [11] consists of down-sizing the majority class by removing observations at random until the dataset is balanced.”