Hands-On Unsupervised Learning with Python

Structure of the dataset

In standard supervised (and often also unsupervised) tasks, the dataset is expected to be balanced. In other words, the number of samples belonging to each class should be almost the same. In the tasks we are going to discuss in this chapter, instead, we assume to have very unbalanced datasets X (containing N samples):

• N_outliers << N, if there is an outlier detection (that is, the dataset is partially dirt;, therefore, we need to find out a way to filter all outliers out)
• N_outliers = 0 (or, more realistically, P(N_outliers > 0) → 0), if there is a novelty detection (that is, we can generally trust the existing samples and focus our attention on the new ones)

The reason for these criteria is quite obvious: let's consider ...