NGBoost: Natural Gradient Boosting for Probabilistic Prediction
This provides a flexible and scalable solution for applications like healthcare and weather forecasting where predictive uncertainty is crucial.
The paper tackles the problem of probabilistic prediction in regression by introducing NGBoost, a gradient boosting algorithm that outputs full probability distributions for uncertainty estimation, and it matches or exceeds existing methods in performance.
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. An open-source implementation is available at github.com/stanfordmlgroup/ngboost.