* **This version represents a major change from the last release (v0.6), which was released one year and half ago.**
* Updated Sklearn API
- Add compatibility layer for scikit-learn v0.18: `sklearn.cross_validation` now deprecated
- Updated to allow use of all XGBoost parameters via `**kwargs`.
- Updated `nthread` to `n_jobs` and `seed` to `random_state` (as per Sklearn convention); `nthread` and `seed` are now marked as deprecated
- Updated to allow choice of Booster (`gbtree`, `gblinear`, or `dart`)
- `XGBRegressor` now supports instance weights (specify `sample_weight` parameter)
- Pass `n_jobs` parameter to the `DMatrix` constructor
- Add `xgb_model` parameter to `fit` method, to allow continuation of training
* Refactored gbm to allow more friendly cache strategy
- Specialized some prediction routine
* Robust `DMatrix` construction from a sparse matrix
* Faster construction of `DMatrix` from 2D NumPy matrices: elide copies, use of multiple threads
* Automatically remove nan from input data when it is sparse.
- This can solve some of user reported problem of istart != hist.size
* Fix the single-instance prediction function to obtain correct predictions
* Minor fixes
- Thread local variable is upgraded so it is automatically freed at thread exit.
- Fix saving and loading `count::poisson` models
- Fix CalcDCG to use base-2 logarithm
- Messages are now written to stderr instead of stdout
- Keep built-in evaluations while using customized evaluation functions
- Use `bst_float` consistently to minimize type conversion
- Copy the base margin when slicing `DMatrix`
- Evaluation metrics are now saved to the model file
- Use `int32_t` explicitly when serializing version
- In distributed training, synchronize the number of features after loading a data matrix.
* Migrate to C++11
- The current master version now requires C++11 enabled compiled(g++4.8 or higher)
* Predictor interface was factored out (in a manner similar to the updater interface).
* Makefile support for Solaris and ARM
* Test code coverage using Codecov
* Add CPP tests
* Add `Dockerfile` and `Jenkinsfile` to support continuous integration for GPU code
* New functionality
- Ability to adjust tree model's statistics to a new dataset without changing tree structures.
- Ability to extract feature contributions from individual predictions, as described in [here](http://blog.datadive.net/interpreting-random-forests/) and [here](https://arxiv.org/abs/1706.06060).
- Faster, histogram-based tree algorithm (`tree_method='hist'`) .
- GPU/CUDA accelerated tree algorithms (`tree_method='gpu_hist'` or `'gpu_exact'`), including the GPU-based predictor.
- Monotonic constraints: when other features are fixed, force the prediction to be monotonic increasing with respect to a certain specified feature.
- Faster gradient calculation using AVX SIMD
- Ability to export models in JSON format
- Support for Tweedie regression
- Additional dropout options for DART: binomial+1, epsilon
- Ability to update an existing model in-place: this is useful for many applications, such as determining feature importance
* Python package:
- New parameters:
- `learning_rates` in `cv()`
- `shuffle` in `mknfold()`
- `max_features` and `show_values` in `plot_importance()`
- `sample_weight` in `XGBRegressor.fit()`
- Support binary wheel builds
- Fix `MultiIndex` detection to support Pandas 0.21.0 and higher
- Support metrics and evaluation sets whose names contain `-`
- Support feature maps when plotting trees
- Compatibility fix for Python 2.6
- Call `print_evaluation` callback at last iteration
- Use appropriate integer types when calling native code, to prevent truncation and memory error
- Fix shared library loading on Mac OS X
* R package:
- New parameters:
- `silent` in `xgb.DMatrix()`
- `use_int_id` in `xgb.model.dt.tree()`
- `predcontrib` in `predict()`
- `monotone_constraints` in `xgb.train()`
- Default value of the `save_period` parameter in `xgboost()` changed to NULL (consistent with `xgb.train()`).
- It's possible to custom-build the R package with GPU acceleration support.
- Enable JVM build for Mac OS X and Windows
- Integration with AppVeyor CI
- Improved safety for garbage collection
- Store numeric attributes with higher precision
- Easier installation for devel version
- Improved `xgb.plot.tree()`
- Various minor fixes to improve user experience and robustness
- Register native code to pass CRAN check
- Updated CRAN submission
* JVM packages
- Add Spark pipeline persistence API
- Fix data persistence: loss evaluation on test data had wrongly used caches for training data.
- Clean external cache after training
- Implement early stopping
- Enable training of multiple models by distinguishing stage IDs
- Better Spark integration: support RDD / dataframe / dataset, integrate with Spark ML package
- XGBoost4j now supports ranking task
- Support training with missing data
- Refactor JVM package to separate regression and classification models to be consistent with other machine learning libraries
- Support XGBoost4j compilation on Windows
- Parameter tuning tool
- Publish source code for XGBoost4j to maven local repo
- Scala implementation of the Rabit tracker (drop-in replacement for the Java implementation)
- Better exception handling for the Rabit tracker
- Persist `num_class`, number of classes (for classification task)
- `XGBoostModel` now holds `BoosterParams`
- libxgboost4j is now part of CMake build
- Release `DMatrix` when no longer needed, to conserve memory
- Expose `baseMargin`, to allow initialization of boosting with predictions from an external model
- Support instance weights
- Use `SparkParallelismTracker` to prevent jobs from hanging forever
- Expose train-time evaluation metrics via `XGBoostModel.summary`
- Option to specify `host-ip` explicitly in the Rabit tracker
* Documentation
- Better math notation for gradient boosting
- Updated build instructions for Mac OS X
- Template for GitHub issues
- Add `CITATION` file for citing XGBoost in scientific writing
- Fix dropdown menu in xgboost.readthedocs.io
- Document `updater_seq` parameter
- Style fixes for Python documentation
- Links to additional examples and tutorials
- Clarify installation requirements
* Changes that break backward compatibility
- [1519](https://github.com/dmlc/xgboost/pull/1519) XGBoost-spark no longer contains APIs for DMatrix; use the public booster interface instead.
- [2476](https://github.com/dmlc/xgboost/pull/2476) `XGBoostModel.predict()` now has a different signature