Landmark Isomap ordering (no vegan; avoids O(n^2) dist matrix)

Landmark approximation to Isomap for fast ordering. Builds a weighted kNN graph on rows of X, computes geodesic distances to a subset of landmark points, embeds landmarks by classical MDS, then projects all points by inverse-distance weighted averaging.

isomap_ordering(
  X,
  k = 15,
  ndim = 1,
  component = 1,
  landmark = NULL,
  landmark_method = c("random", "kmeans"),
  seed = NULL,
  keep = c("giant", "all"),
  return_full = TRUE,
  orient_by_pc1 = TRUE,
  scale01 = TRUE,
  eps = 1e-08
)

Arguments

X

numeric matrix (n x D).

k

number of nearest neighbors in the kNN graph (2 <= k < n).

ndim

embedding dimension (>= component).

component

which embedding coordinate to use as ordering.

landmark

integer; number of landmark points. If NULL, uses min(1000, n).

landmark_method

how to choose landmarks: "random" (default) or "kmeans".

seed

optional integer for reproducibility.

keep

component handling: "giant" keeps the largest connected component; "all" uses all nodes (may yield Inf distances if disconnected).

return_full

if TRUE, return length-n t with NA for excluded nodes (keep="giant").

orient_by_pc1

logical; if TRUE, flip sign to align with PC1.

scale01

logical; if TRUE, rescale returned t to [0,1].

eps

small positive number to stabilize inverse-distance weights.

Value

list(t, keep_idx, n_components, embed, landmark_idx, geodesic_to_landmark)