Venue: Proceedings of the 37th International Conference on Machine Learning
Year: 2020
Paper: https://arxiv.org/abs/1910.03225
Abstract
We present Natural Gradient Boosting (NGBoost), an algorithm for generic probabilistic prediction via gradient boosting. Typical regression models return a point estimate, conditional on covariates, but probabilistic regression models output a full probability distribution over the outcome space, conditional on the covariates. This allows for predictive uncertainty estimation – crucial in applications like healthcare and weather forecasting. NGBoost generalizes gradient boosting to probabilistic regression by treating the parameters of the conditional distribution as targets for a multiparameter boosting algorithm. Furthermore, we show how the Natural Gradient is required to correct the training dynamics of our multiparameter boosting approach. NGBoost can be used with any base learner, any family of distributions with continuous parameters, and any scoring rule. NGBoost matches or exceeds the performance of existing methods for probabilistic prediction while offering additional benefits in flexibility, scalability, and usability.
Additional information
Further information: https://stanfordmlgroup.github.io/projects/ngboost/
Github: https://github.com/stanfordmlgroup/ngboost
User guide: https://stanfordmlgroup.github.io/ngboost/intro.html