Comparing anomaly detection and imbalanced classification

For detecting positive cases against negative cases, the standard go-to family of methods would be classification. For the problems described, as long as you have historical data on at least a few positive and negative cases, you can use classification algorithms. However, you have a very common problem: there are only very few observations that are anomalies. This is a problem that is generally known as the problem of imbalanced data.

The problem of imbalanced data

Imbalanced datasets are datasets in which the target class has very unevenly distributed occurrences. An often-occurring example is website sales: among 1,000 visitors, you often have at least 900 visitors that are just watching and browsing, as opposed to maybe 100 who actually buy something.

Using classification methods carelessly on imbalanced data is prone to errors. Imagine that you fit a classification model that needs to predict for each website visitor...