v21.12.00

🚨 Breaking Changes

• Fix indexing of PCA to use safer types (#4255) @lowener
• RF: Add Gamma and Inverse Gaussian loss criteria (#4216) @venkywonka
• update RF docs (#4138) @venkywonka

🐛 Bug Fixes

• Update conda recipe to have explicit libcusolver (#4392) @dantegd
• Restore FIL convention of inlining code (#4366) @levsnv
• Fix SVR intercept AttributeError (#4358) @lowener
• Fix is_stable_build logic for CI scripts (#4350) @ajschmidt8
• Temporarily disable rmm devicebuffer in array.py (#4333) @dantegd
• Fix categorical test in python (#4326) @levsnv
• Revert "Merge pull request #4319 from AyodeAwe/branch-21.12" (#4325) @ajschmidt8
• Preserve indexing in methods when applied to DataFrame and Series objects (#4317) @dantegd
• Fix potential CUDA context poison when negative (invalid) categories provided to FIL model (#4314) @levsnv
• Using sparse expanded distances where possible (#4310) @cjnolet
• Fix for mean_squared_error (#4287) @viclafargue
• Fix for Categorical Naive Bayes sparse handling (#4277) @lowener
• Throw an explicit excpetion if the input array is empty in DBSCAN.fit #4273 (#4275) @viktorkovesd
• Fix KernelExplainer returning TypeError for certain input (#4272) @Nanthini10
• Remove most warnings from pytest suite (#4196) @dantegd

📖 Documentation

• Add experimental GPUTreeSHAP to API doc (#4398) @hcho3
• Fix GLM typo on device/host pointer (#4320) @lowener
• update RF docs (#4138) @venkywonka

🚀 New Features

• Add GPUTreeSHAP to cuML explainer module (experimental) (#4351) @hcho3
• Enable training single GPU cuML models using Dask DataFrames and Series (#4300) @ChrisJar
• LinearSVM using QN solvers (#4268) @achirkin
• Add support for exogenous variables to ARIMA (#4221) @Nyrio
• Use opt-in shared memory carveout for FIL (#3759) @levsnv
• Symbolic Regression/Classification C/C++ (#3638) @vimarsh6739

🛠️ Improvements

• Fix Changelog Merge Conflicts for branch-21.12 (#4393) @ajschmidt8
• Pin max dask and distributed to 2012.11.2 (#4390) @galipremsagar
• Fix forward merge #4349 (#4374) @dantegd
• Upgrade clang to 11.1.0 (#4372) @galipremsagar
• Update clang-format version in docs; allow unanchored version string (#4365) @zbjornson
• Add CUDA 11.5 developer environment (#4364) @dantegd
• Fix aliasing violation in t-SNE (#4363) @zbjornson
• Promote FITSNE from experimental (#4361) @lowener
• Fix unnecessary f32/f64 conversions in t-SNE KL calc (#4331) @zbjornson
• Update rapids-cmake version (#4330) @dantegd
• rapids-cmake version update to 21.12 (#4327) @dantegd
• Use compute-sanitizer instead of cuda-memcheck (#4324) @teju85
• Ability to pass fp64 type to cuml benchmarks (#4323) @teju85
• Split treelite fil import from forest object definition (#4306) @levsnv
• update xgboost version (#4301) @msadang
• Accounting for RAFT updates to matrix, stats, and random implementations in detail (#4294) @divyegala
• Update cudf matrix calls for to_numpy and to_cupy (#4293) @dantegd
• Update conda recipes for Enhanced Compatibility effort (#4288) @ajschmidt8
• Increase parallelism from 4 to 8 jobs in CI (#4286) @dantegd
• RAFT distance prims public API update (#4280) @cjnolet
• Update to UCX-Py 0.23 (#4274) @pentschev
• In FIL, clip blocks_per_sm to one wave instead of asserting (#4271) @levsnv
• Update of "Gracefully accept 'n_jobs', a common sklearn parameter, in NearestNeighbors Estimator" (#4267) @NV-jpt
• Improve numerical stability of the Kalman filter for ARIMA (#4259) @Nyrio
• Fix indexing of PCA to use safer types (#4255) @lowener
• Change calculation of ARIMA confidence intervals (#4248) @Nyrio
• Unpin dask & distributed in CI (#4235) @galipremsagar
• RF: Add Gamma and Inverse Gaussian loss criteria (#4216) @venkywonka
• Exposing KL divergence in TSNE (#4208) @viclafargue
• Unify template parameter dispatch for FIL inference and shared memory footprint estimation (#4013) @levsnv