weird 3.2.0
Breaking changes
glosh_scores() and lof_scores() now use
the k nearest neighbours excluding the observation itself.
Previously the observation itself was also included, so scores will
differ from earlier versions.
lof_scores() no longer replaces infinite values by
1.
New features
- New data sets:
fashion (Fashion-MNIST sneakers with
planted images of other classes), road_deaths (every road
crash death in Australia from 1989 to 2025), us_mortality
(US weekly deaths and death rates, 2015–2025), and
us_senate_edges and us_senate_members (US
Senate co-voting networks from the 40th to the 113th Congress, with the
name, state, party and ICPSR identifier of each senator).
fetch_oz_books() downloads the text of 29 Oz books by
L. Frank Baum and Ruth Plumly Thompson.
fetch_pets2006() downloads frames from the PETS2006
video.
surprisals() and surprisals_prob() now
have methods for factors, giving the surprisals and exact surprisal
probabilities of categorical data. The category probabilities are
estimated from the data (optionally with a Dirichlet prior
added to each count, and with loo = TRUE for leave-one-out
estimates), or supplied with prob.
Minor improvements and bug
fixes
kde_bandwidth(method = "lookout") no longer crashes R
when the data contain points that differ only by floating-point error
(e.g. PCA scores of identical observations); mlpack::emst()
could overflow the C stack on such near-duplicates.
surprisals() and surprisals_prob() are
much faster for Gaussian mixtures (e.g. from
dist_mclust()), which are now evaluated directly rather
than one observation at a time. They also give finite surprisals for
observations far into the tails, where the mixture density previously
underflowed to zero.
surprisals_prob() now fits the GPD approximation to
rescaled excesses, so the optimiser converges when the surprisals above
the threshold are large. Previously the fit could stop at the iteration
limit, giving inaccurate tail probabilities. A warning is now given if
the fit does not converge.
weird 3.1.0
dist_kde() HDR calculations now always use the density
at the data points (previously this required at least 200 observations),
following hdrcde.
dist_density() removed in favour of
distributional::dist_density().
gg_density() now labels the axis with the variable name
when plotting a dist_kde object, if available.
hdr_regions() added, showing which highest density
region each observation falls in, for one or more coverage
probabilities, from a univariate or bivariate dist_kde
object.
kde_bandwidth() gains method = "scv", a
smoothed cross-validation bandwidth estimator using
ks::hscv()/ks::Hscv().
hampel_anomalies() gains alpha and
approximation arguments, allowing anomalies to be declared
using a significance level rather than a fixed number of standard
deviations.
weird 3.0.0
- Moved the Old Faithful Geyser examples from the README into a new
vignette,
vignette("old-faithful").
augment() method added for rrcov Pca*
objects, returning the principal component scores, score distances and
orthogonal distances, optionally alongside the original data.
dist_mclust() added to convert a Gaussian mixture model
(Mclust object) to a distributional object.
density_df() added, converting a distributional object
to a data frame containing density values over a grid, suitable for
plotting.
biplot_projection() added, drawing a two-dimensional
projection of scores with the original variable axes overlaid as
arrows.
outlier_map() added, drawing the score distance against
the orthogonal distance for a prcomp or rrcov
Pca* object, with optional cutoff thresholds classifying
observations as regular, good leverage points, orthogonal outliers or
bad leverage points.
- Added
fetch_air_quality() to download Beijing air
quality data from 12 monitoring stations (2013-2017).
- Added
gun_deaths dataset containing firearm homicide
rates and gun ownership by country for 2017.
wine_reviews columns have been reordered.
gg_density_layer() removed,
lof_scores() now returns 1 (instead of 0) when there
are too many duplicates.
mvscale() now better handles missing and infinite
values, and returns centering and scaling terms as attributes.
dist_kde() is now robust to missing values.
- Removed several dependencies
- Bug fixes and documentation improvements
weird 2.1.0
- Added
surprisals() and surprisals_prob()
for glm objects
- Faster calculation of
surprisals_prob() for
lm and gam objects
- More accurate rank-based surprisal calculation when ties and
missings present
- More accurate HDR calculations
- Added covariance calculation for multivariate
dist_kde
objects
- Added
distributional::hdr() and
distributional::parameters() methods for
dist_kde objects
- Faster calculation for
kde_bandwidth() with
method = "lookout", now using mlpack for merge distances
mvscale() re-added to weird
- Dropped lookout
package dependency
- Added lots of unit tests
- Better docs
- Bug fixes
weird 2.0.0
- Added
fr_mortality data set
- Updated
oldfaithful and cricket_batting
data sets
- Refactored package to use distributional
objects
- Added
dist_kde() and dist_density()
- Removed
as_kde() and autoplot.kde()
- Rewrote
kde_bandwidth() to handle more methods
- Added
gg_density() and
gg_density_layer()
- Replaced
density_scores() with
surprisals() and added surprisals_prob()
- Updated
gg_bagplot() and gg_hdrboxplot()
to use show_points argument
- Removed
lookout()
- Added
hampel_anomalies()
mvscale() moved to lookout package
- Dropped
interpolation dependency
- No longer exporting
weird_packages() and
weird_conflicts()
weird 1.0.2
- Removed
wine_reviews dataset and created
fetch_wine_reviews() function.
- Bug fixes.
weird 1.0.0