v21.10.00

🚨 Breaking Changes

• RF: python api behaviour refactor (#4207) @venkywonka
• Implement vector leaf for random forest (#4191) @RAMitchell
• Random forest refactoring (#4166) @RAMitchell
• RF: Add Poisson deviance impurity criterion (#4156) @venkywonka
• avoid paramsSolver::{n_rows,n_cols} shadowing their base class counterparts (#4130) @yitao-li
• Apply modifications to account for RAFT changes (#4077) @viclafargue

🐛 Bug Fixes

• Update scikit-learn version in conda dev envs to 0.24 (#4241) @dantegd
• Using pinned host memory for Random Forest and DBSCAN (#4215) @divyegala
• Make sure we keep the rapids-cmake and cuml cal version in sync (#4213) @robertmaynard
• Add thrust_create_target to install export in CMakeLists (#4209) @dantegd
• Change the error type to match sklearn. (#4198) @achirkin
• Fixing remaining hdbscan bug (#4179) @cjnolet
• Fix for cuDF changes to cudf.core (#4168) @dantegd
• Fixing UMAP reproducibility pytest failures in 11.4 by using random init for now (#4152) @cjnolet
• avoid paramsSolver::{n_rows,n_cols} shadowing their base class counterparts (#4130) @yitao-li
• Use the new RAPIDS.cmake to fetch rapids-cmake (#4102) @robertmaynard

📖 Documentation

• Expose train_test_split in API doc (#4234) @hcho3
• Adding docs for .get_feature_names() inside TfidfVectorizer (#4226) @mayankanand007
• Removing experimental flag from hdbscan description in docs (#4211) @cjnolet
• updated build instructions (#4200) @shaneding
• Forward-merge branch-21.08 to branch-21.10 (#4171) @jakirkham

🚀 New Features

• Experimental option to build libcuml++ only with FIL (#4225) @dantegd
• FIL to import categorical models from treelite (#4173) @levsnv
• Add hamming, jensen-shannon, kl-divergence, correlation and russellrao distance metrics (#4155) @mdoijade
• Add Categorical Naive Bayes (#4150) @lowener
• FIL to infer categorical forests and generate them in C++ tests (#4092) @levsnv
• Add Gaussian Naive Bayes (#4079) @lowener
• ARIMA - Add support for missing observations and padding (#4058) @Nyrio

🛠️ Improvements

• Pin max dask and distributed versions to 2021.09.1 (#4229) @galipremsagar
• Fea/umap refine (#4228) @AjayThorve
• Upgrade Treelite to 2.1.0 (#4220) @hcho3
• Add option to clone RAFT even if it is in the environment (#4217) @dantegd
• RF: python api behaviour refactor (#4207) @venkywonka
• Pytest updates for Scikit-learn 0.24 (#4205) @dantegd
• Faster glm ols-via-eigendecomposition algorithm (#4201) @achirkin
• Implement vector leaf for random forest (#4191) @RAMitchell
• Refactor kmeans sampling code (#4190) @Nanthini10
• Gracefully accept 'n_jobs', a common sklearn parameter, in NearestNeighbors Estimator (#4178) @NV-jpt
• Update with rapids cmake new features (#4175) @robertmaynard
• Update to UCX-Py 0.22 (#4174) @pentschev
• Random forest refactoring (#4166) @RAMitchell
• Fix log level for dask tree_reduce (#4163) @lowener
• Add CUDA 11.4 development environment (#4160) @dantegd
• RF: Add Poisson deviance impurity criterion (#4156) @venkywonka
• Split FIL infer_k into phases to speed up compilation (when a patch is applied) (#4148) @levsnv
• RF node queue rewrite (#4125) @RAMitchell
• Remove max version pin for dask & distributed on development branch (#4118) @galipremsagar
• Correct name of a cmake function in get_spdlog.cmake (#4106) @robertmaynard
• Apply modifications to account for RAFT changes (#4077) @viclafargue
• Warnings are errors (#4075) @harrism
• ENH Replace gpuci_conda_retry with gpuci_mamba_retry (#4065) @dillon-cullinan
• Changes to NearestNeighbors to call 2d random ball cover (#4003) @cjnolet
• support space in workspace (#3752) @jolorunyomi