OCDocker.OCScore.Analysis.NNUtils package

This module provides utility functions for analyzing autoencoder (AE)-based feature importance and neural network representations, including permutation importance evaluation.

Usage:

import OCDocker.OCScore.Analysis.NNUtils as ocnnutils

OCDocker.OCScore.Analysis.NNUtils.run_ae_feature_importance(ae_model, X_valid, y_valid, features, n_repeats=30, save_dir='plots', prefix='AE')[source]

Run permutation-based feature importance on an AE + NN pipeline.

Parameters:

  • ae_model (torch.nn.Module) – Trained model that outputs predictions from latent representations.

  • X_valid (np.ndarray) – Validation feature matrix (already encoded if AE is separate).

  • y_valid (np.ndarray) – Ground-truth RMSE or AUC values.

  • features (list[str]) – Names of original features.

  • n_repeats (int) – Number of permutation repetitions.

  • save_dir (str) – Directory to save the barplot output.

  • prefix (str) – Filename prefix for saved figure.

Returns:

Sorted DataFrame with feature importances.

Return type:

pd.DataFrame