Chapter 24: Imbalanced Learning

We define the phenomenon of imbalanced data sets and explain its consequences on accuarcy. Furthermore, we introduce some techniques for handling imbalanced data sets.

We introduce performance measures other than accuracy and explain their advantages over accuracy for imbalanced date. In addition we introduce extensions of these measures for multiclass settings.

We introduce the concept of a Cost Matrix, the Minimum expected cost priciple and the optimal theoretical threshold.

In this section we focus on empirical thresholding and model-agnostic Meta Costs.

We explain the concepts of instance specific costs and cost-sensitive OVO.

We introduce cost curves for misclassif error and explain the duality between ROC points and cost lines.

We explain cost curves with cost matrices and comparing classifiers. In addition we do a wrap-up comparision to ROC.

We introduce the idea of sampling methods for dealing with imbalanced data. In addition, we explain certain undersampling techniques.

We introduce the state-of-art oversampling technique SMOTE.