NGBoost

Simple NGBoost (Natural Gradient Boosting for Probabilistic Prediction) Implementation.

• Original: https://github.com/stanfordmlgroup/ngboost
• paper: https://arxiv.org/abs/1910.03225

Quick Start

Install Docker and docker-compose on your local machine.

docker-compose up -d --build

docker exec -it ngboost-jupyter bash

Run

python ngboost.py

Conditions

• Data: Artificial Data
  • Input: 1-dimension
  • Target: true_function + random noise scaled by input X
• NGBoost Parameters:
  • Weak Learner: Decision Tree
  • Distribution: Normal Distribution
    • re-parameterized \theta = ln (var (i.e. \sigma^2))
  • Use Natural Gradient
  • n_estimators: 500
  • learning_rate 1e-2

Result