Cluster the units of a scale into a new frame — cluster

Groups the units at frame by the similarity of their profiles and returns the scale with those groups added as a new, coarser frame.

cluster_scale(
  data,
  x,
  k,
  frame = NULL,
  key = NULL,
  value = NULL,
  method = c("pam", "hclust", "kmeans"),
  distance = "euclidean",
  scale_units = FALSE,
  new_frame = "cluster",
  labels = "c",
  hclust_method = "ward.D2",
  seed = NULL,
  ...
)

Arguments

data: Long data: unit codes, one or more identifier columns saying which observation each row is, and a value column. A unit that is not observed over the same set as the others is an error rather than a hole.
x: A NestedScale.
k: Number of clusters.
frame: Frame whose units are clustered. Defaults to the atom frame.
key, value: Columns of data; inferred when unambiguous.
method: "pam" (partitioning around medoids, via the cluster package), "hclust" (stats::hclust) or "kmeans" (stats::kmeans). "pam" and "hclust" work from the distance matrix, so they honour distance=; "kmeans" works on the raw profiles and ignores it.
distance: A name from list_scale_distances().
scale_units: Standardise each unit's profile before measuring; see scale_distance().
new_frame: Name for the frame the clusters form.
labels: How to name the clusters: a prefix (the default "c" gives c01, c02, ...), or "medoid" to name each cluster after its representative unit, which keeps the codes meaningful.
hclust_method: Linkage for method = "hclust".
seed: Seed for "kmeans", which starts from a random assignment. NULL leaves the session's stream alone.
...: Passed to the underlying algorithm.

Value

The scale, with new_frame inserted directly above frame. The assignment is also returned as the "clustering" attribute: a data frame of unit, cluster and (for medoid-based methods) whether the unit is its cluster's medoid.

Details

The new frame goes directly above the clustered one, which is the only position that is always true: the clusters are coarser than frame by construction, but they need not nest inside any existing coarser frame, and nestedscales does not require them to. Use scale_nests() to ask whether a given pair happens to nest.

Examples

s <- scale_example()
set.seed(1)
d <- data.frame(
  unit = rep(scale_units(s), each = 4),
  t = rep(sprintf("t%d", 1:4), 7),
  v = as.numeric(seq_len(28))
)
cl <- cluster_scale(d, s, k = 3)
scale_frames(cl)
#> [1] "sector"  "class"   "group"   "cluster" "unit"   
head(attr(cl, "clustering"))
#>   unit cluster medoid
#> 1   U1     c01  FALSE
#> 2   U2     c01   TRUE
#> 3   U3     c01  FALSE
#> 4   U4     c02   TRUE
#> 5   U5     c02  FALSE
#> 6   U6     c03  FALSE