1.8.2

Breaking changes

None.

New features

• Added support for One-Model-Per-Target (OMPT)-style multi-target models:
  • Regression
  • Binary classification

See https://github.com/jpmml/jpmml-xgboost/issues/70

• Improved support for random forest-style models.

XGBoost boosting and bagging model files are structurally identical.

The two model types can be told apart from one another by observing the value of the /learner/gradient_booster/gbtree/model/gbtree_model_param/num_parallel_tree integer attribute. Boosting models have it set to 1, whereas bagging models (eg. random forests) have it set to the number of member decision trees models in the ensemble model,

All the member decision tree models are stored in a gbtree/model/trees array attribute. When dealing with OMPT-style multi-target models, then the model converter must partition the contents of this attribute into multiple sub-arrays, one for each target.

Currently, the partitioning algorithm is a little shaky. Integration tests confirm its correctness for "pure boosting" and "pure bagging" sub-types. However, it may yield incorrect results for the mixed "boosted bagging" sub-type.

• Improved support for early stopping.

If the training process was halted using the early stopping mechanism, then this is indicated by setting the best_iteration and best_score attributes on the native booster object.

XGBoost 1.X versions tended to duplicate this information at the Python wrapper object level. For example, in the form of a xgboost.core.Booster.best_ntree_limit attribute.

XGBoost 2.X versions no longer do this. The converter must now check for the early stopping status by loading the native booster object and querying its /learner/attributes/best_iteration and /learner/attributes/best_score attributes.

Minor improvements and fixes

• Updated integration testing resources to XGBoost 2.0.3