NEWS.md

aorsf 0.1.5 (unreleased)

• fixed an issue where omitting NA values would cause an error in regression forests.

aorsf 0.1.4

• orsf_vs now returns a column that contains non-reference coded variable names (see #52).

• orsf_vs no longer throws an error when n_predictor_min = 1 is used (see #58).

• orsf_summarize_uni now allows specification of a class to summarize for oblique classification forests (see #57).

• fixed an issue where orsf would throw an uninformative error when all predictors were categorical (see #56)

• oblique random forests can now compute out-of-bag predictions on modified versions of their training data (see #54)

• Setting oobag_pred_type to 'none' when growing a forest no longer necessitates the specification of pred_type when calling predict later (see #48).

• Setting sample_fraction to 1 will no longer result in empty oobag_rows in the forest object (this would cause R to crash when the forest was passed to C++; see #48)

• Re-worked the creation and maintenance of oobag_denom in C++ routines (see #48).

• Restricted mean survival time is now used for pred_type = 'time' instead of median survival time (See #46).

aorsf 0.1.3

• minor changes to partial dependence vignette to resolve code sanitization errors.

aorsf 0.1.2

• Allowed option "time" for pred_type in predict and partial dependence to predict survival time (see #37).

• Added pred_spec_auto() for more convenient specification of variables for partial dependence.

• Partial dependence now runs much faster with multiple threads.

• Added orsf_vint() to compute variable interaction scores using partial dependence.

• Added orsf_update(), which can copy and modify an obliqueForest or modify it in place.

• Added orsf_control functions for classification, regression, and survival (#25).

• optimization implemented for matrix multiplication during prediction (#20)

aorsf 0.1.1

• Fixed an uninitialized value for pd_type

• Fixed various issues related to memory leaks

aorsf 0.1.0

• Re-worked internal C++ routines following the design of ranger.

• Re-worked how progress is printed to console when verbose_progress is TRUE, following the design of ranger. Messages now indicate the action being taken, the % complete, and the approximate time until finishing the action.

• Improved variable importance, following the design of ranger. Importance is now computed tree-by-tree instead of by aggregate. Additionally, mortality is the type of prediction used for importance with survival trees, since mortality does not depend on pred_horizon.

• Allowed multi-threading to be performed in orsf(), predict.orsf_fit(), and functions in the orsf_vi() and orsf_pd() family.

• Allowed sampling without replacement and sampling a specific fraction of observations in orsf()

• Included Harrell's C-statistic as an option for assessing goodness of splits while growing trees.

• Fixed an issue where an uninformative error message would occur when pred_horizon was > max(time) for orsf_summarize_uni. Thanks to @JyHao1 and @DustinMLong for finding this!

aorsf 0.0.7

• Additional changes in internal testing to avoid problems with ATLAS

aorsf 0.0.6

• Minor fix for internal tests that were failing when run on ATLAS

aorsf 0.0.5

• orsf() no longer throws errors or warnings when you try to give it a single predictor. A note was added to the documentation in the details of ?orsf that explains why using a single predictor with orsf() is somewhat useless. This was done to resolve mlr-org/mlr3extralearners#259.

• predict.orsf_fit now accepts pred_horizon = 0 and returns sensible values. Thanks to @mattwarkentin for the feature request.

• added a function to perform variable selection, orsf_vs().

• Made variable importance consistent with respect to group_factors. Originally, the output from orsf would have ungrouped VI values while orsf_vi would have grouped values. With this update, orsf defaults to grouped values. The ungrouped values can still be recovered.

• Fixed an issue in orsf_pd functions where output data were not being returned on the original scale.

aorsf 0.0.4

• orsf formulas now accepts Surv objects (see #11)

• Added verbose_progress input to orsf, which prints messages to console indicating progress.

• Allowance of missing values for orsf. Mean and mode imputation is performed for observations with missing data. These values can also be used to impute new data with missing values.

• Centering and scaling of predictors is now done prior to growing the forest.

aorsf 0.0.3

• Included rOpenSci reviewers Christopher Jackson, Marvin N Wright, and Lukas Burk in DESCRIPTION as reviewers. Thank you!

• Added clarification to docs about pros/cons of different variable importance techniques

• Added regression tests for aorsf versus obliqueRSF (they should be similar)

• Additional support and tests for functions with long right hand sides

• Updated out-of-bag vignette with more appropriate custom functions.

• Allow status values in input data to be more general, i.e., not just 0 and 1.

• Allow missing values in predict functions, including partial dependence.

aorsf 0.0.2

• Modified unit tests for compatibility with extra checks run through CRAN.

aorsf 0.0.1

• Added orsf_control_custom(), which allows users to submit custom functions for identifying linear combinations of inputs while growing oblique decision trees.

• Added weights input to orsf, allowing users to over or under fit orsf to specific data in their training set.

• Added chf and mort options to predict.orsf_fit(). Mortality predictions are not fully implemented yet - they are not supported in partial dependence or out-of-bag error estimates. These features will be added in a future update.

aorsf 0.0.0.9000

• Core features implemented: fit, interpret, and predict using oblique random survival forests.

• Vignettes + Readme covering usage of core features.

• Website hosted through GitHub pages, managed with pkgdown.