time-robust-forest 0.1.4

Explores time information to train a robust random forest

time-robust-forest

A Proof of concept model that explores timestamp information to train a random forest with better Out of Distribution generalization power.

Installation

pip install -U time-robust-forest

How to use it

There are a classifier and a regressor under time_robust_forest.models. They follow the sklearn interface, which means you can quickly fit and use a model:

from time_robust_forest.models import TimeForestClassifier

features = ["x_1", "x_2"]
time_column = "periods"
target = "y"

model.fit(training_data[features + [time_column]], training_data[target])
predictions = model.predict_proba(test_data[features])[:, 1]

There are only a few arguments that differ from a traditional Random Forest. two arguments

• time_column: the column from the input dataframe containing the time periods the model will iterate over to find the best splits (default: "period")
• min_sample_periods: the number of examples in every period the model needs to keep while it splits.
• period_criterion: how the performance in every period is going to be aggregated. Options: {"avg": average, "max": maximum, the worst case}. (default: "avg")

Make sure you have a good choice for the time column

Don't simply use a timestamp column from the dataset, make it discrete before and guarantee there is a reasonable amount of data points in every period. Example: use year if you have 3+ years of data. Notice the choice to make it discrete becomes a modeling choice you can optimize.

License

This project is licensed under the terms of the BSD-3 license. See LICENSE for more details.

Citation

@misc{time-robust-forest,
  author = {Moneda, Luis},
  title = {Time Robust Forest model},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/lgmoneda/time-robust-forest}}
}