Hands-On Unsupervised Learning with Python

Anomalies as outliers or novelties

The main topic of this chapter is the automatic detection of anomalies without any supervision. As the models are not based on feedback provided by labeled samples, we can only rely on the properties of the whole dataset to find out the similarities and highlight the dissimilarities. In particular, we start from a very simple but effective assumption: common events are normal, while unlikely events are generally treated as anomalies. Of course, this definition implies that the process we are monitoring is working properly and the majority of outcomes are considered as valid. For example: a silicon-processing factory has to cut a wafer into equal chunks. We know that each of them is 0.2 × 0.2 inches (about ...