README.md

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 = TimeForestClassifier(time_column=time_column)

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

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

• time_column: the column from the input data frame containing the 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 model will aggregate the performance in every period. Options: {"avg": average, "max": maximum, the worst case}. (default: "avg")

To use the environment-wise optimization:

from time_robust_forest.hyper_opt import env_wise_hyper_opt

params_grid = {"n_estimators": [30, 60, 120],
              "max_depth": [5, 10],
              "min_impurity_decrease": [1e-1, 1e-3, 0],
              "min_sample_periods": [5, 10, 30],
              "period_criterion": ["max", "avg"]}

model = TimeForestClassifier(time_column=time_column)

opt_param = env_wise_hyper_opt(training_data[features + [time_column]],
                               training_data[TARGET],
                               model,
                               time_column,
                               params_grid,
                               cv=5,
                               scorer=make_scorer(roc_auc_score,
                                                  needs_proba=True))

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 are a reasonable number of data points in every period. For example, use year if you have 3+ years of data. Notice that the choice to make it discrete becomes a modeling choice you can optimize.

Random segments

Selecting randomly from multiple time columns

The user can use a list instead of a string as the time_column argument. The model will select randomly from it when building every estimator from the defined n_estimators.

from time_robust_forest.models import TimeForestClassifier

features = ["x_1", "x_2"]
time_columns = ["periods", "periods_2"]
target = "y"

model = TimeForestClassifier(time_column=time_columns)

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

Generating random segments from a timestamp column

The user can define a maximum number of segments (random_segments), and the model will split the data using the time stamp information. In the following example, the model segments the data into 1, 2, 3, and 10 parts. For every estimator, it randomly picks one of the ten columns representing the time_column and uses it. In this case, the time_column should be the time stamp information.

from time_robust_forest.models import TimeForestClassifier

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

model = TimeForestClassifier(time_column=time_column, random_segments=10)

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

License

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

Useful links

• Introducing the Time Robust Tree blog post
• Paper

Citation

@inproceedings{moneda2022time,
  title={Time Robust Trees: Using Temporal Invariance to Improve Generalization},
  author={Moneda, Luis and Mauá, Denis},
  booktitle={Brazilian Conference on Intelligent Systems},
  pages={385--397},
  year={2022},
  organization={Springer}
}