Consensus Multidimensional Scaling (CoMDS)

Usage

coMDS(
  D_list,
  embed_list = NULL,
  ndim = 2,
  constraint = c("indscal", "idioscal", "identity"),
  ...
)

Arguments

D_list: List of distance matrices.
embed_list: Optional list of embeddings to compute distances from. If provided, the pairwise Euclidean distances will be computed between the embedded samples and used in lieu of D_list.
ndim: Number of dimensions for the output embedding. Default is 2.
constraint: Type of weight constraint to use in the CoMDS algorithm. Options are "indscal" (default), "idioscal", or "identity"
...: Additional arguments passed to the smacofIndDiff() function.

Value

A list containing the following components:

gspace: Consensus embedding matrix of size n x ndim, where n is the number of samples.
cweights: A list of learned weight matrices for each input embedding/ distance matrix.
stress: The final stress value of the CoMDS solution.
spp: Stress per point (in percent).
rss: Residual sum of squares.
spps: Stress per point per input embedding/distance matrix (in percent, conditional on the input embedding/distance matrix).
sps: Stress per input embedding/distance matrix (in percent).
sps_norm: Stress per input embedding/distance matrix (in percent), normalized by sum of source weights.
ndim: Number of dimensions in consensus embedding.
model: Type of smacof model.
niter: Number of iterations run.
nobj: Number of points/samples.
type: MDS type used (e.g., "interval", "ratio", "ordinal", or "mspline").
constraint: Type of weight constraint used in the CoMDS algorithm.
call: The matched call to the function.

Examples

data(iris)
# remove duplicates so that tSNE can run
X <- dplyr::distinct(iris[, 1:4])

# fit various dimension reduction methods
pca_scores <- prcomp(X, center = TRUE, scale = TRUE)$x
tsne_scores <- Rtsne::Rtsne(X, dims = 2, perplexity = 30, verbose = FALSE)$Y
umap_scores <- umap::umap(X, n_components = 2, verbose = FALSE)$layout
dr_list <- list(
  pca = pca_scores,
  tsne = tsne_scores,
  umap = umap_scores
)

# fit coMDS using dimension reduction embeddings directly as input
comds_out <- coMDS(embed_list = dr_list, ndim = 2)

# fit coMDS using distance matrices as input
dist_list <- purrr::map(dr_list, ~ dist(.x, method = "euclidean"))
comds_out <- coMDS(D_list = dist_list, ndim = 2)

# can re-run with larger number of iterations or smaller tolerance if desired
comds_out <- coMDS(D_list = dist_list, ndim = 2, itmax = 500, eps = 1e-8)