The problem of imbalanced data

One of the most challenging data issues is imbalanced data, which occurs when one or more class levels are much more common than the others. Many, if not most, machine learning algorithms struggle mightily to learn heavily imbalanced datasets, and although there isn’t a specific threshold that determines when a dataset is too off-balance, the problems caused by the lack of balance become increasingly serious as the problem becomes more severe.

In the early stages of class imbalance, small problems are found. For instance, simple performance measures like accuracy begin to lose relevance and more sophisticated performance measures like those described in Chapter 10, Evaluating Model Performance, are needed. As the imbalance widens, bigger problems occur. For example, with extremely imbalanced datasets, some machine learning algorithms might struggle to predict the minority group at all. With this in mind, it might be wise to begin worrying...