prfr

Probabilistic random forest regressor: random forest model that accounts for errors in predictors and labels, yields calibrated probabilistic predictions, and corrects for bias.

For a faster and more elaborate calibration routine (highly recommended), a JAX installation is required. You can install the package with the extra jax feature, which will install the necessary dependencies.

Installation

From PyPI, with jax feature:

pip install "prfr[jax]" 

From PyPI, without jax feature:

pip install prfr

From Github (latest), with jax feature:

pip install "prfr[jax] @ git+https://github.com/al-jshen/prfr"

From Github (latest), without jax feature:

pip install "git+https://github.com/al-jshen/prfr"

Example usage

import numpy as np
import prfr

x_obs = np.random.uniform(0., 10., size=10000).reshape(-1, 1)
x_err = np.random.exponential(1., size=10000).reshape(-1, 1)
y_obs = np.random.normal(x_obs, x_err).reshape(-1, 1) * 2. + 1.
y_err = np.ones_like(y_obs)

train, test, valid = prfr.split_arrays(x_obs, y_obs, x_err, y_err, test_size=0.2, valid_size=0.2)

model = prfr.ProbabilisticRandomForestRegressor(n_estimators=250, n_jobs=-1)
model.fit(train[0], train[1], eX=train[2], eY=train[3])
model.fit_bias(valid[0], valid[1], eX=valid[2])

# check whether the calibration routine will run with JAX
print(prfr.has_jax)

model.calibrate(valid[0], valid[1], eX=valid[2])

pred = model.predict(test[0], eX=test[2])
pred_qtls = np.quantile(pred, [0.16, 0.5, 0.84], axis=-1)

print(pred.shape)