H2O v3.22.0.2 Release Notes

  • ๐Ÿš€ Download at: http://h2o-release.s3.amazonaws.com/h2o/rel-xia/2/index.html

    Bug

    ๐Ÿ“œ [PUBDEV-3281] - Fixed an issue that caused ARFF parser to parse some file incorrectly. ๐Ÿ›  [PUBDEV-4737] - When performing a grid search in Python, fixed an issue that caused all models to return a model.type of "supervised." [PUBDEV-5352] - When running DRF in the Python client, checkpointing on new data now works correctly. ๐Ÿ›  [PUBDEV-5869] - Fixed an issue that caused the confusion matrix recall and precision values to be switched. ๐Ÿ›  [PUBDEV-6036] - In the Python client, fixed an issue that caused the offset_column parameter to be ignored when it was passed in the GLM train statement. [PUBDEV-6042] - The H2O Tree Handler now works correctly on Isolation Forest models. ๐Ÿ›  [PUBDEV-6046] - When running AutoML, fixed an issue that resulted in a "Failed to get metric: auc from ModelMetrics type BinomialGLM" message. [PUBDEV-6050] - In Flow, Precision and Recall definitions are no longer inverted in the confusion matrix. ๐Ÿ›  [PUBDEV-6052] - Fixed the error message that displays when converting from a pandas dataframe to an h2oframe in Python 3.6. ๐Ÿ›  [PUBDEV-6054] - In XGBoost, fixed an issue that resulted in a "Maximum amount of file descriptors hit" message. ๐Ÿ›  [PUBDEV-6060] - Fixed the description of sample_rate in Isolation Forest. 0๏ธโƒฃ [PUBDEV-6063] - Cross validation models are no longer deleted by default. ๐Ÿ›  [PUBDEV-6065] - When viewing an AutoML leaderboard, fixed an issue that resulted in an ArrayIndexOutOfBoundsException if sort_metric was specified but no model was built.

    New Feature

    [PUBDEV-5766] - Added monotonicity constraints to H2O XGBoost.

    Task

    [PUBDEV-6039] - When generating MOJOs, h2o-genmodel.jar now includes a check for MOJO version 1.3 to determine whether the ho2-genmodel.jar and the MOJO version can work together. Prior versions of h2o-3 did not include MOJO 1.3, and as a result, MOJOs silently returned predicted values executed on an empty vector.

    Improvement

    ๐Ÿ“œ [PUBDEV-5705] - With a new skipped_columns option, users can now specify to drop specific columns before parsing. Note that this functionality is not supported for SVMLight or Avro file formats. [PUBDEV-6062] - The GLM multinomial coefficient table now includes the original levels as column names.

    ๐Ÿ“„ Docs

    ๐ŸŽ [PUBDEV-3216] - Created new Performance & Prediction and Variable Importance sections in the User Guide. 0๏ธโƒฃ [PUBDEV-5313] - Updatd the default value of categorical_encoding for XGBoost. This defaults to Auto (which is one_hot_encoding). โšก๏ธ [PUBDEV-6012] - In the parameter entry for weights_column, updated the example to exclude the weight column in the list of predictors. โšก๏ธ [PUBDEV-6016] - In the DRF FAQ, updated the "What happens when you try to predict on a categorical level not seen during training?" question. ๐Ÿ“„ [PUBDEV-6025] - TargetingEncoder is now included in the Python module docs. ๐Ÿ“š [PUBDEV-6041] - In GLM, updated the documentation to indicate that coordinate_descent is no longer experimental. [PUBDEV-6064] - Added default values for max_depth, sample_size, and sample_rate. Also added a parameter description entry for sample_size, showing an Isolation Forest example. [PUBDEV-6086] - Added the new monotone_constraints option to the XGBoost chapter.