H2O Release 3.46

We are excited to announce the release of H2O-3 3.46.0.1! Some of the highlights of this major release are that we added custom metric support for XGBoost, allowed grid search models to be sorted with custom metrics, and we enabled H2O MOJO and POJO to work with MLFlow. Several improvements were also made to the Uplift model (like MLI support). Another exciting update is that we now allow GLM models that were previously built to be used to calculate full loglikelihood and AIC. We also focused on fixing security vulnerabilities. Please read on for more details!

Security patch updates

H2O should always be deployed behind firewalls and in protected clusters. Many of the reported vulnerabilities assume that H2O isn’t deployed under any protection. Regardless, there are several vulnerabilities we did fix during this release:

• CVE-2023-6016: We introduced a Java Property that disables POJO import (defaults to disabled) to avoid remote code execution (courtesy of Marek Novotný).
• CVE-2023-35116: We upgraded the jackson-databind version to address potential vulnerabilities (thanks to Marek Novotný).
• SYNK-JAVA-CIMNUMBUSDS-6247633: We upgraded the nimbus-jose-jwt version to enhance security and to mitigate potential risks (kudos to Adam Valenta).
• CVE-2023-6038 and CVE-2023-6569: We introduced a new configuration option for filtering file system access during reading and writing in response to security concerns (credit to Bartosz Krasinski).

XGBoost support for customized metrics (Adam Valenta)

XGBoost now supports the custom_metric_func parameter which lets you specify any desired metric. The custom_metric_func parameter is well known from other algorithms like GBM and Deep Learning. To see it in action, please look to our documentation.

Loglikelihood and AIC calculation support for GLM MOJOs (Yuliia Syzon)

GLM MOJOs now support the calculation of loglikelihood and AIC given a dataset with the response column when the MOJO is loaded with the H2O Generic model. To enable this feature, you have to build the GLM model with calc_like=True.

MLFlow support (Marek Novotný)

We added the libraries from MLFlow (and necessary code) to enable you to use H2O-3 MOJO and POJO with the MLFlow frameworks.

Uplift DRF improvements (Veronika Maurerova)

In this release, we added many improvements for the Uplift DRF algorithm to provide you with more opportunities for tuning and interpreting your models. We added grid search, early stopping, partial dependence plots, and variable importance.

Python documentation improvements (Shaun Yogeshwaran)

We improved the Python documentation for the GAM, Model Selection, and ANOVA GLM algorithms by expanding the number of available examples. More will follow in the coming releases!

Contributors

Wendy Wong, Adam Valenta, Marek Novotný, Tomáš Frýda, Veronika Maurerova, Bartosz Krasinski, Yuliia Syzon, Sebastien Poirier, Hannah Tillman, Shaun Yogeshwaran