ENSEMBLE METHODS FOR CLASS IMBALANCE LEARNING - Imbalanced Learning: Foundations, Algorithms, and Applications - page 71

4.4 EMPIRICAL STUDY

Many research works reported the effectiveness of ensemble methods to handle
imbalanced data. To illustrate the advantages of ensemble methods for CIL, we
compare some typical ones with standard methods and some CIL methods without
ensemble techniques. 3
The datasets are 10 binary class-imbalanced University of California, Irvine
(UCI) datasets [39] whose information is summarized in Table 4.1. The methods
in comparison are:

1. Bagging , with CART as base learning algorithm.
2. AdaBoost , with CART as base learning algorithm.
3. RF [19]. RF is a state-of-the-art ensemble method. It injects random-
ness into base learning algorithm instead of training data. Specifically, RF
trains random decision trees as base learners by random feature selection.
When constructing a component decision tree, at each step of split selec-
tion, RF first selects a feature subset randomly, and then carries out the
conventional split selection procedure within the selected feature subset.
4. Under-sampling + AdaBoost. First random under-sampling is used such
that the majority class has the same number of examples as the minority
class, and then an AdaBoost ensemble is trained from the new dataset.
5. Under-sampling + RF. First random under-sampling is used such that
the majority class has the same number of examples as the minority class,
and then a RF ensemble is trained from the new dataset.

Table 4.1 Basic Information of Datasets a
Dataset

Size Attribute Target #min/#maj Ratio

abalone 4177 1N,7C Ring = 7 391/3786 9.7

balance 625 4C Balance 49/576 11.8

cmc 1473 3B,4N,2C class 2 333/1140 3.4

aberman 306 1N,2C class 2 81/225 2.8

housing 506 1B,12C [20, 23] 106/400 3.8

mf-morph 2000 6C class 10 200/1800 9.0

mf-zernike 2000 47C class 10 200/1800 9.0

pima 768 8C class 1 268/500 1.9

vehicle 846 18C opel 212/634 3.0

wpbc 198 33C recur 47/151 3.2

a Size is the number of examples. Targe t is used as the minority class, and all others are used
as the majority class. In Attribute , B :binary, N : nominal, C : continuous. # min /# maj is the size
of minority and majority classes, respectively. Ratio is the size of the majority class divided by
that of the minority class.

3 The results are mainly from [6].