adversarial-labeller 0.1.8

Sklearn compatiable model instance labelling tool to help validate models in situations involving data drift.

adversarial_labeller

---

Adversarial labeller is a sklearn compatible labeller that scores instances as belonging to the test dataset or not to help model selection under data drift. Adversarial labeller is distributed under the MIT license.

Installation

Dependencies

Adversarial validator requires:

• Python (>= 3.7)
• scikit-learn (>= 0.21.0)
• [imbalanced learn](>= 0.5.0)
• [pandas](>= 0.25.0)

User installation

The easiest way to install adversarial validator is using

pip install adversarial_labeller

Example Usage

import numpy as np
import pandas as pd
from sklearn.datasets.samples_generator import make_blobs
from sklearn.metrics import accuracy_score
from sklearn.model_selection import cross_val_score
from sklearn.ensemble import RandomForestClassifier
from adversarial_labeller import AdversarialLabelerFactory, Scorer

scoring_metric = accuracy_score

# Our blob data generation parameters for this example
number_of_samples = 1000
number_of_test_samples = 300

# Generate 1d blob data and label a portion as test data
# ... 1d blob data can be visualized as a rug plot
variables, labels = \
make_blobs(
    n_samples=number_of_samples,
    centers=2,
    n_features=1,
    random_state=0
)

df = pd.DataFrame(
{
    'independent_variable':variables.flatten(),
    'dependent_variable': labels,
    'label': 0  #  default to train data
}
)
test_indices = df.index[-number_of_test_samples:]
train_indices = df.index[:-number_of_test_samples]

df.loc[test_indices,'label'] = 1  # ... now we mark instances that are test data

# Now perturb the test samples to simulate data drift/different test distribution
df.loc[test_indices, "independent_variable"] +=\
np.std(df.independent_variable)

# ... now we have an example of data drift where adversarial labeling can be used to better estimate the actual test accuracy

features_for_labeller = df.independent_variable
labels_for_labeller = df.label

pipeline, flip_binary_predictions =\
    AdversarialLabelerFactory(
        features = features_for_labeller,
        labels = labels_for_labeller,
        run_pipeline = False
    ).fit_with_best_params()

scorer = Scorer(the_scorer=pipeline,
                flip_binary_predictions=flip_binary_predictions)

# Now we evaluate a classifer on training data only, but using
# our fancy adversarial labeller
_X = df.loc[train_indices]\
    .independent_variable\
    .values\
    .reshape(-1,1)

_X_test = df.loc[test_indices]\
            .independent_variable\
            .values\
            .reshape(-1,1)

# ... sklearn wants firmly defined shapes
clf_adver = RandomForestClassifier(n_estimators=100, random_state=1)
adversarial_scores =\
    cross_val_score(
        X=_X,
        y=df.loc[train_indices].dependent_variable,
        estimator=clf_adver,
        scoring=scorer.grade,
        cv=10,
        n_jobs=-1,
        verbose=1)
# ... and we get ~ 0.70 - 0.68
average_adversarial_score =\
    np.array(adversarial_scores).mean()

# ... let's see how this compares with normal cross validation
clf = RandomForestClassifier(n_estimators=100, random_state=1)
scores =\
    cross_val_score(
        X=_X,
        y=df.loc[train_indices].dependent_variable,
        estimator=clf,
        cv=10,
        n_jobs=-1,
        verbose=1)

# ... and we get ~ 0.92
average_score =\
    np.array(scores).mean()

# now let's see how this compares with the actual test score
clf_all = RandomForestClassifier(n_estimators=100, random_state=1)
clf_all.fit(_X,
            df.loc[train_indices].dependent_variable)

# ... actual test score is 0.70
actual_score =\
accuracy_score(
    clf_all.predict(_X_test),
    df.loc[test_indices].dependent_variable
)

adversarial_result = abs(average_adversarial_score - actual_score)
print(f"... adversarial labelled cross validation was {adversarial_result:.2f} points different than actual.")  # ... 0.00 - 0.02 points

cross_val_result = abs(average_score - actual_score)
print(f"... regular validation was {cross_val_result:.2f} points different than actual.")  # ... 0.23 points

#  See tests/ for additional examples, including against the Titanic and stock market trading