Technology Bundle ID
NCI-E-080-2019

A Most Efficient and Convergent Principal Component Analysis (PCA) Method for Big Data

Applications
Co-Inventors
Xiaowei Song (NIDA)
Development Status
Basic (Target Identification)
ICs

Big data usually means big sample size with many outliers, in which traditional scalable L2-norm principal component analysis (L2-PCA) will fail. Current existing L1-norm PCA (L1-PCA) methods can improve robustness over outliers, however, its scalability is usually limited in either sample size or dimension size.  The inventor proposes an online flipping method to solve L1-PCA challenges, which is not only convergent asymptotically (or with big data), but also achieves most efficiency in the sense each sample is visited only once to extract one principal component (PC). The proposed PCA also has certain robustness to outliers compared to L2-PCA.

If you need a linear complexity robust PCA solver, please contact us; our method can even solve robust PCA in real-time. This efficient robust PCA algorithm is available for licensing and/or collaborations to explore utility for your application.

Commercial Applications
  • Big data analysis 
  • This approach may be the indicated procedure in the presence of unbalanced outlier contamination
Competitive Advantages

Current existing L1-norm PCA (L1-PCA) methods can improve robustness over outliers, however, its scalability is usually limited in either sample size or dimension size. The proposed PCA also has certain robustness to outliers compared to L2-PCA

Request More Info

Licensing Contact: