ppforest2 0.1.4

Bug fixes

• For forests trained on classification data whose rows are not sorted by the levels of the response (for example, grouped by class but in a different order), oob_error(), oob_predictions(), permuted_importance() and weighted_importance() used the wrong out-of-bag observations, because each tree’s bootstrap sample referred to the rows after they were sorted for training. The OOB error was too low as a result. The bootstrap samples now refer to the rows in the order given.
• For regression models, model$x, model$y, fitted() and residuals() follow the order of the training rows. They used to be sorted by the response, so fitted() did not line up with the data.
• predict() errors when new_data has a different number of columns from the training data. It used to return predictions computed from mismatched columns.
• predict() errors when new_data contains NA, NaN or infinite values. A data frame row with a missing value used to be dropped silently, so fewer predictions than rows were returned.
• pptr() and pprf() error on missing values in data passed through the formula interface, as they already did for the matrix interface. Rows with missing values used to be dropped silently.
• pptr() and pprf() error on infinite values in the features.
• Class predictions from predict() keep every group of the model as a factor level, including groups that are not predicted.
• stop_min_size(), stop_min_variance() and stop_max_depth() error with a clear message when their argument is not a single value of the right type, instead of R’s “the condition has length > 1” or silently truncating a non-integer.

Improvements

• predict() accepts a data frame without the response column for models trained through the formula interface.
• pptr() and pprf() reject empty data and list columns with explicit messages.
• save_json() appends .json to a path without that suffix, and returns the path it wrote.
• Constant features are detected with a tolerance relative to the feature’s magnitude when scaling variable importance, and summary() detects standardized data with a tolerance.
• A forest’s training specification is converted between C++ and R once per forest instead of once per tree, which makes pprf(), predict() and the other functions that pass a forest to C++ faster for forests of many small trees.

Documentation

• citation("ppforest2") cites the package and the two papers describing the method, and ?ppforest2 documents the package.

Packaging and build

• DESCRIPTION lists the authors of the bundled nlohmann/json and PCG code as copyright holders, and inst/COPYRIGHTS gives their copyright and license.

ppforest2 0.1.3

New features

• summary() on a classification tree or forest now reports a per-class error rate alongside each confusion matrix, and prints the overall error rate above the matrix rather than below it. A class with no observations in the data is shown as - rather than nan% — this happens when the model predicts a class that never appears as an actual label, for example when predicting on a subset of the data.
• Text sizes in the tree structure plot are configurable through options (ppforest2.text_edge, ppforest2.text_tick, ppforest2.text_leaf, ppforest2.text_proj), and ppforest2.text_scale multiplies all of them at once for rendering the plot large.

Bug fixes

• Projection coefficients in the tree structure plot are formatted to three significant digits instead of two fixed decimals, matching the axis tick labels in the same plot. The projector is normalized so the values it projects always have the same spread whatever the units of the input data, which leaves its coefficients at a magnitude set by the data rather than by the split; the iris root projector .00429 / -.0391 / .0259 / .0335 rendered as .00 / .04 / .03 / .03, merging the two petal terms and dropping sepal length to zero.
• pptr() and pprf() no longer abort with the internal error Grouping::init: partition must be rooted at row 0 when the response’s class blocks are contiguous but ordered by decreasing factor level (for example a two-class factor whose first row is its second level, or the bundled crab dataset with default alphabetical levels). The classification path now sorts the response into ascending group-id order whenever it is not already, matching the regression path and the command-line tool.
• The resolved thread count is clamped to at least 1 — hardware_concurrency() may report 0, and a non-positive OpenMP thread count is undefined behavior.

Packaging and build

• The vendored C++ core uses Eigen::indexing::all instead of Eigen::all, which Eigen 5 no longer accepts as an index, so the package builds with the upcoming RcppEigen release based on Eigen 5 as well as with the current one based on Eigen 3.4.0.

ppforest2 0.1.2

CRAN

• Fixed a -Wdeprecated-declarations warning reported by CRAN’s macOS/M1mac additional check (Apple clang 21, macOS 26 SDK). The newer libc++ deprecates std::char_traits<unsigned char>, which the vendored nlohmann/json instantiates through its binary output/stream adapters (std::basic_string<std::uint8_t> / std::basic_ostream<std::uint8_t>). ppforest2 does not use the nlohmann/json binary formats, so the vendored json.hpp is now bracketed with a _Pragma guard that suppresses the deprecation. _Pragma (unlike #pragma) is not flagged by R CMD check’s pragma check. The guard is applied by make r-vendor-deps (scripts/vendor-guard-json.sh).

ppforest2 0.1.1

Packaging and build

• The package now compiles the C++ core directly through Makevars instead of CMake, with no network access or downloaded dependencies at install time. Eigen is provided by RcppEigen; nlohmann/json and pcg headers are vendored under inst/include. This makes the package installable on CRAN’s offline build machines.
• Self-registering strategies are compiled directly into the shared object, removing the previous whole-archive linking workaround.
• Compiler flags mirror the standalone C++ build — EIGEN_NO_AUTOMATIC_RESIZING on all platforms and EIGEN_DONT_VECTORIZE on Windows.
• Replaced C++20 designated initializers and fixed member-initialization order in the vendored core so it compiles warning-free under a strict C++17 GCC (-Wall -Wextra -pedantic).
• A compile-time EIGEN_VERSION_AT_LEAST(3, 4, 0) guard fails the build with a clear message if an incompatible Eigen is supplied via RcppEigen.
• make r-vendor-deps re-vendors the committed json/pcg headers after a version bump.

Documentation

• DESCRIPTION uses Authors@R and cites the projection-pursuit tree and forest references with DOIs.
• Examples for the parsnip and plot methods use \donttest with requireNamespace() guards instead of \dontrun, so they run under --run-donttest when the suggested packages are available.

CRAN

• Added cran-comments.md. The package passes R CMD check --as-cran with no errors or warnings; remaining notes (new submission, cosmetic pragmas in the vendored nlohmann/json headers) are documented for the reviewer.

ppforest2 0.1.0

New features

• Projection-pursuit oblique decision trees and random forests for classification, using LDA/PDA optimization.
• pptr() and pprf() with formula and matrix interfaces. Returned models carry an S3 class vector identifying both model type and mode (e.g. c("pprf_classification", "pprf", "ppmodel")).
• predict() returns group labels (type = "class") or vote proportions (type = "prob") for classification.
• Random uniform variable selection per split for forest diversity.
• Three variable importance measures: permuted (VI1), projections (VI2), and weighted projections (VI3). Permuted variable importance may be negative; this is meaningful signal (“within noise”) rather than a sentinel, so callers should rely on the ranking rather than clipping at zero. Weighted projection importance is non-negative by construction.
• Out-of-bag error and confusion matrix for forests, with bootstrap sample indices persisted for recomputation.
• Lazy OOB accessors — oob_error(), oob_predictions(), oob_samples(), bag_samples(), permuted_importance(), weighted_importance() — compute from the training data stored on the model on first access and memoize in an environment cache, so training is fast and repeated access is free. oob_error() is NA_real_ and oob_predictions() returns a factor with NA for rows with no OOB tree.
• summary() displays training and OOB confusion matrices.
• Degenerate split detection when projection pursuit cannot find a useful projection, surfaced as a warning.
• OpenMP multi-threaded forest training.
• save_json() and load_json() for model persistence. Optional metrics fields use a uniform null-or-value representation so downstream tooling can distinguish “computed but empty” from other shapes without special-casing.
• Cross-platform reproducibility — identical results for the same seed on Linux, macOS, and Windows, enforced by golden-file tests in CI.
• tidymodels/parsnip integration: pp_tree() and pp_rand_forest() model specifications.
• ggplot2 visualizations — tree diagrams, variable importance plots, projection histograms, and decision boundary plots.
• Bundled classification datasets: crab, crabs, fishcatch, glass, image, leukemia, lymphoma, NCI60, olive, parkinson, and wine. (Use datasets::iris from base R for iris examples.)

Experimental features

Regression support is included but untested in production workloads. API surface and defaults may change in future releases.

• Regression auto-detected when y is numeric (not a factor). predict() returns a numeric vector (type = "response").
• Regression strategy wrappers: grouping_by_cutpoint(), leaf_mean_response(), stop_min_size(), stop_min_variance(), stop_any(). Training quantile-slices the continuous response into groups and fits mean-response leaves.
• summary() displays MSE / MAE / R² for regression models, computed for training and out-of-bag predictions; forest OOB error is reported as MSE. oob_predictions() returns a numeric vector with NA_real_ for rows with no OOB tree.
• save_json() / load_json() preserve regression mode; parsnip pp_tree() / pp_rand_forest() accept mode = "regression".
• Bundled regression dataset california_housing (20,433 × 9, predict median_house_value). For smaller regression examples use datasets::mtcars from base R.