In this chapter, we examined anomalies in data. We discussed several approaches to anomaly detection and looked at two kinds of anomalies: outliers and novelties. We considered the fact that anomaly detection is primarily an unsupervised learning problem, but despite this, some algorithms require labeled data, while others are semi-supervised. The reason for this is that, generally, there is a tiny number of positive examples (that is, anomalous samples) and a large number of negative examples (that is, standard samples) in anomaly detection tasks.

In other words, we usually don't have enough positive samples to train algorithms. That is why some solutions use labeled data to improve algorithm generalization and precision. On the contrary, supervised learning usually requires a large number of positive and negative examples, and their distribution needs to be balanced.

Also, notice that the task of detecting anomalies does not have a single formulation and that it is often interpreted differently, depending on the nature of the data and the goal of the concrete task. Moreover, choosing the correct anomaly detection method depends primarily on the task, data, and available a priori information. We also learned that different libraries can give slightly different results, even for the same algorithms.

In the following chapter, we will discuss the dimension reduction methods. Such methods help us to reduce the dimensionality of data with high dimensionality into a new representation of data with lower dimensionality while preserving the essential information from the original data.