rcccd 0.4.0

Speed and memory

PCCCD cover: the greedy dominant-ball search is rewritten in C++ and computes distances on the fly, so the n x n distance and indicator matrices are no longer built. Memory use is now linear in the number of samples.
RWCCCD cover: the sorting of distances is done once per sample instead of in every iteration of the greedy search; only the weights change between iterations. This removes most of the computation time on larger data.
predict() methods compute the nearest-ball ratios in C++ instead of building a full dominant-by-new-sample distance matrix with ‘proxy’. The ‘proxy’ dependency is dropped, and the ‘RcppArmadillo’ dependency is no longer needed.

rwcccd_classifier(): the T score used an integer division (n_uncovered/2) in the C++ code, so it differed from rwcccd_classifier_2() whenever the number of uncovered samples was odd. Both functions now give the same result. Dominant balls (and therefore predictions with the default e = 0) are unchanged in most cases, but can differ in a few data sets because of this fix.
A predictor matrix with a single column or a class with a single observation no longer fails (x[i, ] dropped the dimensions).
Ties in predict() are now broken deterministically (first class) instead of randomly.
Classes without any observation in the training data are allowed (e.g. in a bootstrap sample of the ensemble); they get probability zero.
pcccd_ensemble_classifier() resamples when a sample contains a single class, and its predict() is vectorized.

Input checks for all classifiers: x must be numeric without missing values, y must be a factor of matching length without missing values, and at least two classes must be observed.
rwcccd_classifier_2() is kept as a slow pure R reference implementation.
Added tests (testthat), including checks of the C++ covers against plain R implementations.
f_cover_pcccd() (internal, not exported) keeps its old signature so that packages calling it, such as ‘imbalanceDatRel’, keep working.