Some Notes on Clustering
Available Clustering Options
Any clustering method that comes from the Scikit-learn clustering module
Shared Nearest Neighbor clustering (implemented by us inside of the
cfl.clustering_methodssubmodule)Any other module that uses the same interface as a Scikit-learn clusterer
Recommendations for Choosing a Clusterer
DBSCAN and KMeans are the two clustering methods we’ve worked with the most, so unless you have a reason to choose another, maybe stick with one of those.
KMeans - Advantages: only one parameter to tune, meaning of parameter (# of clusters) is pretty intuitive - Potential Disadvantages: can only detect globular clusters, forces the user to choose the number of clusters (a goal of CFL is to detect number of macrovariables without supervision)
DBSCAN - Advantages: does not force you to pre-define number of clusters, can detect clusters of any shape, you can maybe get away with only tuning one parameter (eps) - Disadvantages: has two parameters (eps and min_samples) (even though eps is more important to tune than min_samples), can be tricky to tune well, may not correctly distinguish two clusters if there is overlap between the clusters
Shared Nearest Neighbor - a derivative of DBSCAN designed to perform well on high dimensional data
Optuna is one of various tools that you can use to select clustering hyperparameters.