News

Changes in 3.0:

• Parallel training with OpenMP.
• New organisation of the internal PRNG.
• Error propagation reporting removed (arguments reportErrorEvery and saveErrorPropagation) since it brings too little value for the complexity it requires. And can be emulated with merging anyway.

Changes in 2.0.3:

• Registration of native routines.
• Documentation updates.

Changes in 2.0.2:

• Added a wrapper for performing feature selection using shadow importance.

Changes in 2.0.1:

• Fixed bug in merge.rFerns code, causing merged importance scores to be wrong, especially after merging very small ensembles.
• Fixed memory protection bug triggering rare segmentation faults in the R's GC, often long after the rFerns model was completed.

Changes in 2.0.0:

• Shadow importance.
• Roxygen documentation.
• Model merging for parallel, batch and deeply traceable training.
• Fixed a minor error in leaf score calculation (basically redundant to the model performance). Thanks to Mateusz Fedoryszak for spotting this.
• Fixed OOB predictions returned by predict(), which are now always data.frame, even in the multi-label case.

Changes in 1.1.0:

• Code portability fixes.
• predict method returns OOB results for no input data even when fern forest is not saved.
• Fixed minor importance calculation bug.

Changes in 1.0.0:

• Published in JSS.

Changes in 0.3.3:

• Fixed uninitialised memory bug causing problems with LLVM.

Changes in 0.3.2:

• Prototype of a multi-label classification.
• Removed redundant constant from scores definition; score values will be different, but the importance scores and prediction results will be the same as generated by previous versions.

Changes in 0.3.1:

• predict method for rFerns returns OOB votes/scores when no predictors are given.

Changes in 0.3:

• Handling of unbalanced classes.
• Performance tweaks.

Changes in 0.2:

• Added support for integer and ordered-factor attributes, suggested by Bill Venables.

Changes in 0.1:

• First public release.

Reference manual