Hands-On Machine Learning for Algorithmic Trading

Hierarchical DBSCAN

Hierarchical DBSCAN is a more recent development that assumes clusters are islands of potentially differing density, to overcome the DBSCAN challenges just mentioned. It also aims to identify the core and non-core samples. It uses the min_cluster_ size and min_samples parameters to select a neighborhood and extend a cluster. The algorithm iterates over multiple eps values and chooses the most stable clustering.

In addition to identifying clusters of varying density, it provides insight into the density and hierarchical structure of the data.

The following screenshots show how DBSCAN and HDBSCAN are able to identify very differently shaped clusters: