riverreliability 0.1.2

The river reliability diagram provides insight into classifier calibration with a visualization and metric based on the posterior balanaced accuracy.

River reliability

Install

Install the package with:

pip install riverreliability

How to use

Below, we show some basic funtionality of the package. Please look at the notebooks for more examples and documentation.

np.random.seed(42)

We start of by generating a fake dataset for classification and splitting it in a train and test set.

X, y = sklearn.datasets.make_classification(n_samples=5000, n_features=12, n_informative=3, n_classes=3)
X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(X, y, test_size=0.2, shuffle=True)

For this example we use an SVM. We fit it on the training data and generate probabilities for the test set.

logreg = sklearn.svm.SVC(probability=True)
logreg.fit(X_train, y_train)
y_probs = logreg.predict_proba(X_test)

As a sanity check we compute some performance metrics.

print(f"Accuracy: {sklearn.metrics.accuracy_score(y_test, y_probs.argmax(axis=1))}")
print(f"Balanced accuracy: {sklearn.metrics.balanced_accuracy_score(y_test, y_probs.argmax(axis=1))}")

Accuracy: 0.808
Balanced accuracy: 0.8084048918146675

To get an insight into calibration we can look at the posterior reliability diagrams and the PEACE metric.

We can plot the diagrams aggregated over all classes:

ax = riverreliability.plots.river_reliability_diagram(y_probs.max(axis=1), y_probs.argmax(axis=1), y_test, bins="fd")

peace_metric = riverreliability.metrics.peace(y_probs.max(axis=1), y_probs.argmax(axis=1), y_test)
ax.set_title(f"PEACE: {peace_metric:.4f}")

_ = ax.legend()

Or class-wise to spot miscalibrations for particular classes:

import matplotlib.pyplot as plt

axes = riverreliability.plots.class_wise_river_reliability_diagram(y_probs, y_probs.argmax(axis=1), y_test, bins=15)

peace_metric = riverreliability.metrics.class_wise_error(y_probs, y_probs.argmax(axis=1), y_test, base_error=riverreliability.metrics.peace)
_ = plt.suptitle(f"PEACE: {peace_metric:.4f}")

In this particular example we can see that the classifier is well calibrated.

See the notebooks directory for more examples.