Testing

Comparison with the R implementation

To ensure functional equivalence with the original R implementation of ICS, ICSpyLab is validated against the reference package through automated tests.

The testing workflow consists of the following steps:

1. Setup: Install necessary dependencies, including pytest and rpy2 for interfacing with R.

2. Data Loading: Load datasets such as iris, wine, and diabetes using scikit-learn.

3. ICS execution: Perform the ICS (Invariant Coordinate Selection) algorithm in both R (using rpy2) and Python.

4. Validation: Outputs from the Python implementation are compared against those obtained from the R package to assess numerical consistency.

Fixtures and Parameters

To streamline the testing process, fixtures are used to load data and run the ICS algorithm in both R and Python. Parameters for the ICS algorithm, such as covariance estimators and transformation settings, are defined and tested across different datasets and configurations.

Validation

The results with algorithms ‘standard’ and ‘whiten’ from the Python implementation are validated against the R implementation by comparing:

• Transformation matrices

• Kurtosis values

• Skewness values (if available)

• Transformed data

Comparisons are performed using numerical tolerance thresholds to account for floating-point implementation differences between R and Python. The method is numpy.testing.assert_almost_equal with 7 decimals. This ensures that icspylab reproduces the behavior of the reference implementation with a high level of consistency.

Unit tests

Additional unit tests cover:

• validation of the ICS estimator against scikit-learn estimator checks (sklearn.utils.estimator_checks.check_estimator),

• object initialization and parameter validation,

• error handling and invalid input scenarios,

• consistency between ‘standard’ and ‘whiten’ algorithms,

• affine invariance properties of the transformed coordinates.

The complete testing suite is available in the tests directory of the source repository.