Almost Linear-Time k-Medoids Clustering

Introduction

banditpam is an R package that lets you do \(k\)-mediods clustering efficiently as described in Tiwari, et. al. (2020).

We illustrate with a simple example using simulated data from a Gaussian Mixture Model with the the following means: \((0, 0)\), \((-5, 5)\) and \((5, 5)\).

set.seed(10)
n_per_cluster <- 40
means <- list(c(0, 0), c(-5, 5), c(5, 5))
X <- do.call(rbind, lapply(means, MASS::mvrnorm, n = n_per_cluster, Sigma = diag(2)))

Let’s cluster the observations in this X matrix using 3 clusters. The first step is to create a KMedoids object:

obj <- KMedoids$new(k = 3)

Next we fit the data with a specified loss, \(l_2\) here. A good habit is to set the seed before fitting for reproducibility.

set.seed(198)
obj$fit(data = X, loss = "l2")

And we can now extract the medoid observation indices.

med_indices <- obj$get_medoids_final()

A plot shows the results where we color the medoids in red.

d <- as.data.frame(X); names(d) <- c("x", "y")
dd <- d[med_indices, ]
ggplot(data = d) +
  geom_point(aes(x, y)) +
  geom_point(aes(x, y), data = dd, color = "red")
Clustering with 3-mediods with L2 loss
Clustering with 3-mediods with L2 loss

We can also change the loss function and see how the mediods change.

obj$fit(data = X, loss = "l1")  # L1 loss
med_indices <- obj$get_medoids_final()
Clustering with 3-mediods with L1 loss
Clustering with 3-mediods with L1 loss

One can query some performance statistics too; see help on KMedoids.

obj$get_statistic("dist_computations") # no of dist computations
#> [1] 32517
obj$get_statistic("cache_misses") #  no of cache misses
#> [1] 0

References

Tiwari, Mo, Martin J Zhang, James Mayclin, Sebastian Thrun, Chris Piech, and Ilan Shomorony. 2020. “BanditPAM: Almost Linear Time k-Medoids Clustering via Multi-Armed Bandits.” In Advances in Neural Information Processing Systems, 368–74.