v0.2.10

• Added the branch difference method for calculating contributions.
• Normalized logloss metric, by diving it by the sun of the sample weight.

v0.2.9

Hot fix to add default values for GOSS parameters to booster.

v0.2.8

This release add the GOSS (Gradient Based One Side Sampling) sampling method to the package. Additionally small performance improvements, and moved the application of the learning_rate till after the monotonicity bounds are generated.

v0.2.7

This release introduces the ability to record metrics on data while fitting, by passing in data to the evaluation_data parameter. Additionally early_stopping_rounds are now supported so training will be cut short, if there is no improvement seen in performance for a specified number of iterations.

v0.2.6

Made the partial dependence predictions faster, and improvements to python API.

v0.2.5

This release adds the method parameter to the predict_contributions metric. This allows either "average" (the method xgboost uses to calculate the contribution matrix when approx_contribs is True), or "weight" to be specified. The "average" method averages the leaf weights across all nodes, and then uses this for calculating the contributions. The "weight" method instead uses the internal leaf weights, to determine how a feature impacts the final score. Both methods result in a contribution matrix, that if summed by each row, is equivalent to the output of the predict method.

v0.2.4

This release adds an experimental parameter, create_missing_branch, that if True, will create a separate branch for missing, creating a ternary tree, the missing node will be given the same weight value as the parent node. If this parameter is False, missing will be sent down either the left or right branch, creating a binary tree. Defaults to False.

Release v0.2.3

• Fixed bug where features were said to have zero variance, with low populated levels.
• Added subsample parameter to sample records prior to training each tree.
• Added missing parameter, so missing can now be assigned to any float, rather than only being np.nan.
• Added metadata store to booster object, so info can be saved on the booster and loaded again later after saving.

v0.2.1

This release adds the predict_contributions method to the GradientBooster object, both for the Rust and Python API.

v.0.2.2

Hot fix to a bug in the contributions prediction where the base_score would be ignored. Additionally, performance improvements were made, to make contributions calculations ~5X faster.