nirs4all.visualization.analysis.transfer module

class nirs4all.visualization.analysis.transfer.PreprocPCAEvaluator(r_components=10, knn=10)

Bases: object

fit(raw_data: dict[str, ndarray], pp_data: dict[str, dict[str, ndarray]])

raw_data: {“dataset”: X_raw_(n,m), …} pp_data: Can be either:

• {“pp_name”: {“dataset”: X_pp_(n,p), …}, …} OR

• {“dataset”: {“pp_name”: X_pp_(n,p), …}, …}

(will automatically detect and pivot if needed)

Assumes rows (samples) are aligned within each dataset across raw and pp.

get_cross_dataset_summary(metric='centroid_improvement')

Get a summary of how preprocessing affects inter-dataset distances.

Parameters:

metric – ‘centroid_improvement’ or ‘spread_improvement’ Higher values = preprocessing brought datasets closer

Returns:

DataFrame sorted by improvement (best preprocessing first)

get_quality_metric_convergence()

Analyze how preprocessing affects the similarity of quality metrics across datasets. Lower variance = preprocessing makes datasets more homogeneous in quality.

Returns:

DataFrame with variance of quality metrics (evr, cka, rv, etc.) across datasets for raw vs preprocessed data. Lower values = better convergence.

plot_all(show=True)

Generate all visualization plots.

plot_all_datasets_pca(figsize=(16, 12))

Plot all datasets together in the same PCA space for raw and each preprocessing. Shows how datasets cluster and separate in different preprocessing spaces.

plot_cross_dataset_distances(figsize=(14, 8))

Plot how preprocessing affects inter-dataset distances. Shows which preprocessing methods bring datasets closer together.

plot_cross_dataset_heatmap(metric='centroid_improvement', figsize=(12, 10))

Create a heatmap showing pairwise dataset distances for each preprocessing.

Parameters:

metric – ‘centroid_improvement’, ‘centroid_dist_pp’, ‘spread_improvement’, or ‘spread_dist_pp’

plot_distance_matrices(metric='centroid', figsize=(18, 12))

Plot distance matrices showing inter-dataset distances for raw and all preprocessings. Shows which preprocessing reduces distances (better for transfer learning).

Parameters:

metric – ‘centroid’ or ‘spread’ - which distance metric to display

plot_distance_reduction_ranking(metric='centroid', log_scale=False, figsize=(14, 8))

Bar chart showing which preprocessing methods best reduce inter-dataset distances. Directly answers: “Which preprocessing is best for transfer learning?”

Parameters:

• metric – ‘centroid’ or ‘spread’ - which distance method to use for ranking

• log_scale – If True, use log scale for the right plot (absolute distances) to handle extreme values

plot_pair(dataset: str, preproc: str, figsize=(10, 5))

Enhanced comparison plot for a specific dataset-preprocessing pair.

plot_preservation_summary(by='preproc', figsize=(14, 8))

Enhanced summary plot with better styling.

plot_quality_metric_convergence(figsize=(16, 10))

Visualize how preprocessing makes quality metrics more homogeneous across datasets. Shows variance reduction in EVR, CKA, RV, Procrustes, Trustworthiness, Grassmann.

Lower variance after preprocessing = datasets behave more similarly = better for transfer learning.