fit_model() now treats numeric or character class labels as classes; previously a numeric response (such as Y in data69_1) was fitted as a regression.
fit_model() only checks the columns used in the formula, so missing values in unrelated columns no longer stop fitting.
boundary_compute() can recompute a multi-model result, and recomputing a boundary no longer repeats the "classbound_boundary" class.
plot_boundary(type = "disagreement") no longer fails when obs_data is supplied.
boundary_explore_nd() rescales features before its k-nearest-neighbour fallback for models without class probabilities, as classifly does; previously one large-scale feature could dominate and boundary points were close to random. It also keeps non-syntactic feature names and accepts character columns.
fit_model() evaluates transformations in the formula (such as log(x) or I(x^2)) with the formula interface; previously the classifier failed with “object not found”. The features of the model are the original variables.
fit_model() now stops with a clear message when the response has fewer than two classes, instead of failing inside the classifier.
plot_boundary() with obs_data and classbound() accept a projection basis without row names (such as one taken directly from tourr), assuming the rows follow the order of the training features, instead of failing.
classbound() takes the axes from the fitted model, so formulas containing transformations work.
print() on a classbound object prints the fitted model on separate lines instead of a literal \n.
The internal tidymodels_bridge() now passes resolution on.
Shiny application
In Projection mode, the projection plane is recomputed when new data are generated; previously the old plane was kept, which broke the plots when the number of dimensions changed.
Imported data keep rows whose missing values are only in columns the app does not use, and data with fewer than two numeric features are rejected up front.
Custom models can declare supports_prob = TRUE, and fitted parsnip models from the workspace can show a probability surface.
Clear Canvas also resets the undo history and the number of base classes.
Exported grid predictions and the reproduce script now use the same plot range as the app, and the script passes on the prediction arguments (such as the PPtree rule). Each export uses its own temporary folder.
The export wizard creates its ZIP file with the ‘zip’ package when it is installed, so exporting no longer needs a zip program (often missing on Windows).
The reproduce script writes zoom limits with full precision, and the exported data.csv has an is_outlier column so the script redraws the outliers as diamonds, as in the app.
A zero-size brush no longer causes a division by zero when zooming.
An empty or invalid Point Density value no longer stops the app while drawing.
Outliers belonging to a class that has been removed no longer cause ordinary observations to be drawn as outliers.
Only installed models are preselected at start-up.
The information table reports the 2D Slice mode correctly.
New features
lda and qda models from ‘MASS’ are supported out of the box, with their posterior probabilities available for the probability surface.
Documentation
Projection views are now described as what they show: the predicted regions on a plane through the data center (a planar section), with the observations projected onto that plane.
boundary_explore_nd() credits the ‘classifly’ package and Wickham, Cook and Hofmann (2015).
The data69_1 help page now states that Y is numeric with values 0, 1, 2 and that the data are the full UCI dataset.
The README and the getting-started vignette show installation from CRAN, and the README has a CRAN status badge.
Tests
Added regression tests for all of the fixes above (the suite now has 96 tests), tests for classbound(), and tests of the Shiny application’s server logic with shiny::testServer().