Statistical Assertions

The ml_assert.stats module provides functions for statistical comparison and drift detection between datasets.

---

Quick Start

from ml_assert.stats.drift import assert_no_drift

# Check for drift between reference and current datasets
assert_no_drift(df_ref, df_cur, alpha=0.05)

---

Drift Detection

High-Level Drift Detection

The assert_no_drift function automatically detects drift in all columns of a DataFrame.

from ml_assert.stats.drift import assert_no_drift

# Check for drift between training and test sets
assert_no_drift(df_train, df_test, alpha=0.05)

Parameters: - df_ref: Reference DataFrame (e.g., training data) - df_cur: Current DataFrame (e.g., inference data) - alpha: Significance level for statistical tests (default: 0.05)

Low-Level Statistical Tests

Kolmogorov-Smirnov Test

For comparing continuous distributions.

from ml_assert.stats.drift import ks_test

# Compare two numeric samples
statistic, p_value = ks_test(sample1, sample2, alpha=0.05)

Chi-Squared Test

For comparing categorical distributions.

from ml_assert.stats.drift import chi2_test

# Compare two categorical samples
statistic, p_value = chi2_test(sample1, sample2, alpha=0.05)

Wasserstein Distance

For measuring the distance between distributions.

from ml_assert.stats.drift import wasserstein_distance

# Calculate Wasserstein distance
distance = wasserstein_distance(sample1, sample2)

---

Distribution Assertions

Distribution Testing

Assert that a dataset follows a specific distribution.

from ml_assert.stats.distribution import assert_distribution

# Assert normal distribution
assert_distribution(data, distribution="normal", alpha=0.05)

Supported Distributions: - "normal" - "uniform" - "exponential" - "poisson"

---

Examples

Basic Drift Detection

from ml_assert.stats.drift import assert_no_drift

# Check for drift in all columns
assert_no_drift(df_train, df_test, alpha=0.05)

Custom Statistical Tests

from ml_assert.stats.drift import ks_test, chi2_test

# Compare numeric columns
statistic, p_value = ks_test(df_train["age"], df_test["age"], alpha=0.05)

# Compare categorical columns
statistic, p_value = chi2_test(df_train["category"], df_test["category"], alpha=0.05)

Distribution Testing

from ml_assert.stats.distribution import assert_distribution

# Test for normal distribution
assert_distribution(data, distribution="normal", alpha=0.05)

# Test for uniform distribution
assert_distribution(data, distribution="uniform", alpha=0.05)

For more detailed API reference, see Stats API.