Skip to contents

HDBSCAN Clustering

Source: R/hdbscan.R
shoal_hdbscan.Rd

Hierarchical density-based spatial clustering of applications with noise.

Usage

shoal_hdbscan(
  x,
  alpha = 1,
  min_samples = 15L,
  min_cluster_size = 15L,
  metric = c("euclidean", "cosine"),
  boruvka = TRUE,
  partial_labels = NULL
)

Arguments

x

A numeric matrix or data frame. Data frames are coerced to a matrix using their numeric columns (non-numeric columns are dropped).

alpha

Smoothing parameter for mutual reachability distance. Default 1.0.

min_samples

Minimum neighbourhood size. Default 15L.

min_cluster_size

Minimum cluster size. Default 15L.

metric

Distance metric, one of "euclidean" or "cosine".

boruvka

Whether to build the minimum spanning tree with a tree-accelerated Boruvka search rather than Prim's algorithm. Default TRUE, which is the faster choice in low dimensions. The acceleration depends on a spatial index, and above a few dozen columns it becomes slower than the plain search: on 20,000 rows in 64 dimensions Boruvka takes around 2.5 times as long as boruvka = FALSE. Both give the same clustering.

partial_labels

Optional named list for semi-supervised clustering. Names are cluster IDs (as strings), values are integer vectors of 1-indexed point indices. NULL (default) for fully unsupervised clustering.

Value

An object of class c("shoal_hdbscan", "shoal_clustering"): a list with components cluster (integer vector of cluster IDs, NA for noise), n_clusters, n_noise, data (the input matrix), algorithm, params, metric, and outlier_scores (GLOSH scores).

Examples

res <- shoal_hdbscan(as.matrix(iris[, 1:4]))
res
#> 
#> ── HDBSCAN Clustering 
#> Metric: "euclidean"
#> Parameters: alpha = 1, min_samples = 15, min_cluster_size = 15, boruvka = TRUE
#> Clusters: 2, Noise points: 0
#> GLOSH outlier scores: median 0.159, max 0.657