ENSEMBLE METHODS FOR CLASS IMBALANCE LEARNING - Imbalanced Learning: Foundations, Algorithms, and Applications - page 73

Table 4.2 The Methods in Comparison
Methods

Abbr. Type

1 Bagging Standard ensemble method

Bagg

Group 1 2 AdaBoost Standard ensemble method

Ada

3 Random Forests RF Standard ensemble method

4 Under-sampling Under-Ada CIL method

+ AdaBoost

5 Under-sampling Under-RF CIL method

+ Random Forests

Group 2 6 Over-sampling Over-Ada CIL method

+ AdaBoost

7 Over-sampling CIL method

Over-RF

+ Random Forests

8 S OTE CIL method

SMOTE

+ under-sampling

+ AdaBoost

9 Chan + AdaBoost Bagging-style method for CIL

Chan

10 Balanced Bagging-style method for CIL

BRF

Random Forests

Group 3 11 AsymBoost Boosting-based method for CIL

Asym

12 SMOTEBoost SMB Boosting-based method for CIL

13 EasyEnsemble Easy Hybrid ensemble method for CIL

14 BalanceCascade Cascade Hybrid ensemble method for CIL

Table 4.3 summarizes the results of area under the curve (AUC) values [40]
by conducting 10 times 10-fold cross-validation. AUC is a popular performance
measure in CIL. The higher the AUC value, the better the performance. Figures
4.1 and 4.2 show the scatter plots of methods in Groups 1 and 2 versus each
of the six ensemble methods for CIL in Group 3, respectively. A point above
the dotted line indicates that the method of y -axes is better than that of the
x -axes. In addition, Table 4.4 gives the detailed win-tie-lose counts of each
class imbalance method (Groups 2 and 3) versus standard methods (Group 1),
and each ensemble method for CIL (Group 3) versus methods in Group 2, via
t -tests with significance level at 0.05. Bold face indicates the result is significant
in sign test at significance level at 0.05.
Almost all the CIL methods have significantly better performance than CART
[6] 5 . But when compared with standard ensemble methods (Group 1), under-
sampling and over-sampling methods in Group 2 are not very effective. It is
probably because standard ensemble methods have strong generalization ability;
this could reduce the effect of class imbalance. While SMOTE is significantly
better than the standard ensemble methods. This suggests that SMOTE sampling
and combination of different sampling methods are good choices to handle imbal-
anced data when invoking ensemble learning algorithm to generate a classifier.

5 Since CART cannot produce AUC values, we did not include CART in comparison list in this
chapter. Liu et al. [6] showed that almost all the CIL methods have significantly higher F -measure
and G -mean values than CART.