Journal of Chemometrics Прикладная математика

Joint Dimension Reduction Method Based on Maximum Entropy

Zhonghai He, Yi Zhang, Haili Huang, Xiaofang Zhang
Wiley · 2026-11-01 · DOI 10.1002/cem.70196

Аннотация

ABSTRACT Assessing sample novelty based on spatial location is a common strategy for determining when to update a calibration model. The high‐dimensional nature of spectral data necessitates reducing it to a two‐dimensional (2D) plane for visual inspection. Existing dimensionality reduction methods primarily aim to preserve the original data structure as much as possible. However, the resulting low‐dimensional representations are often suboptimal for visual inspection, thereby hindering the judgment of model updating timing. To facilitate such update decisions, calibration samples should ideally be uniformly distributed within a region characterized by simple boundaries. Motivated by this, we propose a joint dimensionality reduction method based on maximum entropy. This approach sets entropy as the optimization objective. The entropy is calculated using kernel density estimation combined with spatial binning. The maximum entropy is achieved by iteratively optimizing the projection directions. The 2D scatter plot generated by this joint entropy projection (JEP) exhibits three distinct advantages for visual evaluation: a uniform distribution of calibration samples, regular boundaries, and an equal‐contribution coordinate system. These features enable intuitive visualization of the spatial displacement of new samples relative to the calibration‐sample reference region, providing auxiliary information for applicability‐domain (AD) visualization and model‐updating decisions.

Читать у издателя →