filter_leverage — Hat-matrix or PCA-score leverage filter¶
Group: Sample / feature filters · C ABI: n4m_outlier_detection_high_leverage_*
Description¶
Hat-matrix or PCA score-space leverage filter.
Parameters¶
Name |
Type |
Default |
|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
API and bindings¶
C ABI (ABI 2): n4m_outlier_detection_high_leverage_apply · n4m_outlier_detection_high_leverage_create · n4m_outlier_detection_high_leverage_destroy · n4m_outlier_detection_high_leverage_fit · n4m_outlier_detection_high_leverage_is_fitted · n4m_outlier_detection_high_leverage_threshold. Use the linked public header for the exact signature, configuration, and result handles.
Python (verified public re-export):
from n4m.outlier_detection import HighLeverageFilter
Source signature: HighLeverageFilter(method: str | int = 'hat', threshold_multiplier: float = 2.0, absolute_threshold: float | None = None, n_components: int = 0, center: bool = True).
R: no current source-verified entry point was found for this catalog method.
MATLAB / Octave: no current source-verified entry point was found for this catalog method.
Explanations¶
Bibliographic source¶
Rousseeuw & van Zomeren (1990), Unmasking Multivariate Outliers and Leverage Points, JASA 85, 633–639, DOI 10.1080/01621459.1990.10474920 (https://doi.org/10.1080/01621459.1990.10474920), provides the leverage/outlier context; the fallback and thresholds here are implementation-specific.
Mathematical principle¶
Hat mode computes \(h_i=x_i^T(X^TX+10^{-10}I)^{-1}x_i\) after optional centering. PCA mode computes leverage in retained score space; n_components<=0 selects \(\min(n-1,p,50)\). Unless absolute_threshold in (0,1) is supplied, the cutoff is threshold_multiplier times mean training leverage. Rows at or below the cutoff are kept.
Appropriate uses¶
Detecting calibration samples geometrically remote from the fitted X design space.
Limits and validation¶
When rows are not greater than columns, requested hat mode silently uses PCA. Leverage measures influence geometry, not spectral quality or response error, and thresholds are training-distribution dependent.
Implementation¶
Python role API n4m.outlier_detection.HighLeverageFilter; ABI 2 family n4m_outlier_detection_high_leverage_{create,fit,apply,threshold,destroy}.
The ABI-2 implementation is the n4m_outlier_detection_high_leverage_* lifecycle in libn4m.
Sources and provenance¶
https://github.com/GBeurier/nirs4all-methods/blob/main/cpp/src/core/filters/high_leverage.h; https://doi.org/10.1080/01621459.1990.10474920
See also: methods index · interactive dashboard