YDF is a library to train, evaluate, interpret, and serve Random Forest,
Gradient Boosted Decision Trees, and CART decision forest models.

Getting Started 🧭

A concise and modern API

YDF allows for rapid prototyping and development while minimizing risk of modeling errors.

Deep learning composition

Integrated with Jax, TensorFlow, and Keras.

Cutting-edge algorithms

Include the latest decision forest research to ensure maximum performance.

Source: Yggdrasil Decision Forests: A Fast and Extensible Decision Forests Library (KDD 2023).

Source: A Comparison of Decision Forest Inference Platforms from A Database Perspective (Arxiv 2023).

Fast inference

Compute predictions in a few microseconds. Executed in the tens of millions of times per second in Google.

Installation

To install YDF from PyPI, run:

pip install ydf -U

YDF is available on Python 3.8, 3.9, 3.11 and 3.12, on Windows x86-64, Linux x86-64, and macOS ARM64.

Usage example

import ydf
import pandas as pd

# Load dataset with Pandas
ds_path = "https://raw.githubusercontent.com/google/yggdrasil-decision-forests/main/yggdrasil_decision_forests/test_data/dataset/"
train_ds = pd.read_csv(ds_path + "adult_train.csv")
test_ds = pd.read_csv(ds_path + "adult_test.csv")

# Train a Gradient Boosted Trees model
model = ydf.GradientBoostedTreesLearner(label="income").train(train_ds)

# Look at a model (input features, training logs, structure, etc.)
model.describe()

# Evaluate a model (e.g. roc, accuracy, confusion matrix, confidence intervals)
model.evaluate(test_ds)

# Generate predictions
model.predict(test_ds)

# Analyse a model (e.g. partial dependence plot, variable importance)
model.analyze(test_ds)

# Benchmark the inference speed of a model
model.benchmark(test_ds)

# Save the model
model.save("/tmp/my_model")

Key features

Modeling

• Train Random Forest, Gradient Boosted Trees, Cart, and Isolation Forest models.
• Train classification, regression, ranking, uplifting, and anomaly detection models.
• Plot of decision trees.
• Interpret model (variable importances, partial dependence plots, conditional dependence plots).
• Interpret predictions (counter factual, feature variation).
• Evaluate models (accuracy, AUC, ROC plots, RMSE, confidence intervals, cross-validation).
• Hyper-parameter tune models.
• Consume natively numerical, categorical, boolean, tags, text, and missing values.
• Consume natively Pandas Dataframe, Numpy Arrays, TensorFlow Datasets, CSV files and TensorFlow Records.

Serving

• Benchmark model inference.
• Run models in Python, C++, Go, JavaScript, and CLI.
• Online inference with REST API with TensorFlow Serving and Vertex AI.

Advanced modeling

• Compose YDF models with Neural Network models in TensorFlow, Keras, and JAX.
• Distributed training over billions of examples and hundreds of machines.
• Use cutting-edge learning algorithm such as oblique splits, honest trees, hessian scores, global tree optimizations, optimal categorical splits, categorical-set inputs, dart, extremely randomized trees.
• Apply monotonic constraints.
• Consumes multi-dimensional features.
• Enjoy backward compatibility for model and learners since 2018.
• Edits trees in Python.
• Define custom loss in Python.

Next steps

Read the 🧭 Getting Started tutorial. You will learn how to train a model, interpret it, evaluate it, generate predictions, benchmark its speed, and export it for serving.

Ask us questions on Github. Googlers can join the internal YDF Chat.

Read the TF-DF to YDF Migration guide to convert a TensorFlow Decision Forests pipeline into a YDF pipeline.