CLASS IMBALANCE AND ACTIVE LEARNING - Imbalanced Learning: Foundations, Algorithms, and Applications - page 123

Abalone Breast cancer

100

100

98

99

96

98

94

97

92

96

95 0 90

1000 2000 3000 0 100 200 300 400

Letter Satimage

85

100

99

80

98

97

75

96

70 0

95 0

5000 10,000 15,000 1000 2000 3000 4000

SMOTE Active learning V IRTUAL

Figure 6.8 Comparison of SMOTE, AL, and VIRTUAL on UCI datasets. We present the
g -means (%) ( y -axis) of the current model for the test set versus the number of training
samples ( x -axis) seen.

However, SMOTE converges to higher g -means than AL in some of the cate-
gories, indicating that the virtual positive examples provide additional information
that can be used to improve the model. VIRTUAL converges to the same or
even higher g -means than SMOTE while generating fewer virtual instances.
For the UCI datasets (Fig. 6.8), VIRTUAL performs as well as AL in abalone
in g -means and consistently outperforms AL and SMOTE in the other three
datasets.
In Table 6.4, the support vector imbalance ratios of all the three methods are
lower than the data imbalance ratio, and VIRTUAL achieves the most balanced
ratios of positive and negative support vectors in the Reuters datasets. Despite that
the datasets used have different data distributions, the portion of virtual instances
become support vectors in VIRTUAL consistently and significantly higher than
that in SMOTE. These results confirm the previous discussion that VIRTUAL is
more effective in generating informative virtual instances.