0.4.0

[StatMixedML]

We are excited to announce the release of LightGBMLSS v0.4.0! This release brings a new feature, package updates, stability improvements, and bug fixes. Here are the key highlights of this release:

New Features

Mixture Distributions
LightGBMLSS now supports using mixture distributions for modelling univariate targets! Mixture distributions, or mixture densities, extend the concept of traditional univariate distributions by interpreting the observed data as combinations of multiple underlying processes. In its essence, a mixture distribution is a weighted combination of several component distributions, where each component contributes to the overall mixture distribution. Due to their high flexibility, mixture distributions can portray a diverse range of shapes, making them adaptable to a plethora of datasets. By introducing mixture distributions in LightGBMLSS, users get a better understanding of the conditional distribution of the response variable and achieve a more precise representation of the data generation process.

Stability Improvements

Model Estimation
We have improved the stability of the model estimation process. This results in more consistent and accurate estimation of parameters, leading to better predictions and increased model reliability.

Bug Fixes

In addition to the new features and stability improvements, we have addressed various bugs reported by the community. These bug fixes enhance the overall reliability and usability of LightGBMLSS.

Package Dependency Updates

We have updated some of the package-dependencies to the latest versions.

General

We appreciate the valuable feedback and contributions from our users, which have helped us in making LightGBMLSS even better. We encourage you to update to this latest version to take advantage of the new features and improvements. To get started, check out the documentation and examples.

Thank you for your continued support, and we look forward to your feedback.

Happy modeling!

---

This discussion was created from the release 0.4.0.

[tjjerome]

I'm so excited for mixture distributions! Is there a way to model a mixture distribution using different families for each component? I'm specifically trying to use a compound Poisson and Gaussian model.

[StatMixedML]

Thanks for your interest in the project!

The mixture-distribution is based on PyTorch's MixtureSameFamily, where all component distributions are of the same distribution type, but with different parameterizations. So currently, you cannot combine a Gaussian with a Poission as you suggested. Are you aware of a PyTorch/Pyro implementation of a compound mixture?