Proximum - Statistical Analysis of Ensemble Proximity Matrices
Treats the proximity matrices produced by tree ensembles
as first-class statistical objects rather than as model
by-products. Provides a unified extractor across ensemble
engines, in-bag and out-of-bag definitions, transformations to
dissimilarities with metric diagnostics, corrections that make
an indefinite proximity usable as a kernel, and permutation
inference for comparing two proximity matrices or partitioning
one across the terms of a design. Also provides a Nystrom
approximation, a thresholded sparse representation and a
streaming form that never allocates the matrix at all, for
samples too large to hold it, measures how far the proximity
moves between replicates of the ensemble, and draws each object
through 'ggplot2': the seriated matrix, the configuration it
implies, and the thresholded graph with its communities.
