CLASS IMBALANCE AND ACTIVE LEARNING - Imbalanced Learning: Foundations, Algorithms, and Applications - page 101

+

−

Model

Predictions

Selectively label
examples

Oracle

Unlabeled instance

Figure 6.1 Pool-based active learning.

Feature values

−

+

−

+
+

Training set

Oracle Model

Select instance
for labeling

Unlabeled pool

Figure 6.2 Stream-based active learning.

predictive performance is sufficient, the model may be incorporated into an end
system that feeds the model with unlabeled examples and consumes the predicted
results. A diagram illustrating these two types of AL scenarios is presented in
Figures 6.1 and 6.2, respectively. Owing to the greater attention given to pool-
based AL in recent scientific literature and the additional power available to an
AL system capable of processing a large representation of the problem space at
once, the remainder of this chapter focuses on the latter of these two scenarios,
the pool-based setting.
The most common techniques in AL have focused on selecting examples from
a so-called region of uncertainty, the area nearest to the current model's predic-
tive decision boundary. 1 Incorporating uncertainty into active data acquisition
dates back to research focused on optimal experimental design [2], and has been

1 For instance, the simplest case when performing binary classification would involve choosing
x = argmax x min y P(y | x),y ∈ 0 , 1.