DumME: Mixed Effects Dummy Model
This is an adaptation of MERF (https://github.com/manifoldai/merf). The main difference is that this version is fully compliant with the scikit-learn API.
Other difference include:
- The name: MERF was renamed to the more general MixedEffectsModel
- The default fixed-effects model: dummy model instead of random forest
- The package structure: stripped down to its core and then upgraded to use modern standards
- Test suite: using pytest instead of unittest
Caution
We are currently not maintaining or developing this further. Ideally we would contribute our changes to the original version of MERF (see manifoldai#68). Do reach out if you want to build upon or collaborate with us on this.
Using this version
Install via github:
pip install git+https://github.com/phenology/merfInstantiate the dummy model:
from dumme.dumme import MixedEffectsModel
from dumme.utils import DummeDataGenerator
# Get some sample data
dg = DummeDataGenerator(m=0.6, sigma_b=4.5, sigma_e=1)
df, _ = dg.generate_split_samples([1, 3], [3, 2], [1, 1])
y = df.pop("y")
x = df
# Fit a dummy model
# Notice the signature of the `fit` method: first X and y, and the other args are optional.
me_dummy = MixedEffectsModel()
me_dummy.fit(X, y)
# or
me_dummy.fit(X, y, cluster_column="cluster", fixed_effects=["X_0", "X_1", "X_2"], random_effects=["Z"])
# Predict only accepts X as input. It is assumed new data is structured
# in the same way as the original training data.
new_X = X.copy()
me_dummy.predict(new_X)To get the "original" MERF (but still with the new fit signature):
from sklearn.ensemble import RandomForestRegressor
rf = RandomForestRegressor(n_estimators=300, n_jobs=-1)
me_rf = MixedEffectsModel(rf)
me_rf.fit(X, y)