Density-Based Clustering via Rust • petalcluster

R bindings to the petal-clustering Rust library, providing fast density-based clustering via DBSCAN, HDBSCAN, and OPTICS.

Installation

# install.packages("pak")
pak::pak("belian-earth/petalcluster")

Requires a working Rust toolchain (rustc >= 1.81).

Why density-based clustering?

Classical methods like k-means assume convex, evenly-sized clusters. Real data is rarely that clean. Density-based methods find clusters of arbitrary shape and naturally separate noise.

library(petalcluster)

# -- Cluster with both methods --
km  <- kmeans(rings, centers = 3, nstart = 20)
hdb <- petal_hdbscan(rings, min_cluster_size = 15L, min_samples = 5L)

# -- Side-by-side comparison --
cluster_col <- function(labels) {
  pal <- c("#E41A1C", "#377EB8", "#4DAF4A", "#984EA3")
  cols <- rep("grey60", length(labels))
  valid <- !is.na(labels) & labels > 0
  cols[valid] <- pal[((labels[valid] - 1L) %% length(pal)) + 1L]
  cols
}

cluster_pch <- function(labels) ifelse(is.na(labels), 4L, 1L)

par(mfrow = c(1, 2))
plot(rings, col = cluster_col(km$cluster), main =  "k-means (k=3)")
plot(rings, col = cluster_col(hdb$cluster), pch = cluster_pch(hdb$cluster), main ="HDBSCAN")

Example

R’s built-in quakes dataset contains 1,000 seismic events near Fiji. HDBSCAN finds spatial clusters of earthquake activity without needing to specify the number of groups upfront.

quakes_hdbscan <- petal_hdbscan(
  quakes[, c("lat", "long", "depth", "mag")]
)
quakes_hdbscan
#> 
#> ── HDBSCAN Clustering
#> Metric: "euclidean"
#> Parameters: alpha = 1, min_samples = 15, min_cluster_size = 15
#> Clusters: 8, Noise points: 199

Use xcol and ycol to plot a single pair of variables — here longitude vs latitude reveals the geographic structure of the clusters:

plot(quakes_hdbscan, xcol = "long", ycol = "lat", asp=1)

The default plot produces a scatter plot matrix of all variables:

plot(quakes_hdbscan)

Performance

petalcluster offers a performant alternative for density-based clustering in R, particularly at scale. The plot below compares wall-clock time against the dbscan R package and Python’s scikit-learn, across dataset sizes from 500 to 50,000 points.

Acknowledgements

The heavy lifting is done by the petal-clustering crate, developed and maintained by Petabi petalcluster is simply an R interface to their work.