Can we beat overfitting?—A closer look at Cloarec's PLS algorithm. (7th March 2018)

• Record Type: Journal Article
• Authors: M. Sousa, Pedro F.; Åberg, K. Magnus
• Is Part Of: Journal of chemometrics. Volume 32:Number 6(2018)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2018-03-07
• Subjects: algorithm -- overfitting -- PLS -- partial least squares -- projection to latent structures; Chemistry -- Mathematics -- Periodicals; Chemistry -- Statistical methods -- Periodicals; 542.85
• DOI: 10.1002/cem.3002
• Languages: English
• ISSNs: 0886-9383

Abstract

Abstract: Random noise has been addressed as a cause of overfitting in partial least squares regression. A previous study pinpointed that one of the sources of overfitting resides in the calculation of scores due to the accumulation of noise in the diagonal of the variance‐covariance matrix, and a modified partial least squares regression was proposed with the removal of this diagonal prior to the score calculation. Here, a further modification of the NIPALS algorithm is proposed, with the same ability to overcome overfitting due to noise, but algebraically more similar to the original NIPALS. The results indicate that it is possible to get more reliable auto‐prediction R 2 with a cross‐validation performance close to that of the original NIPALS algorithm. Abstract : The NIPALS algorithm has been modified by Cloarec to reduce overfitting to noise inX . Here, a further modification achieving better prediction performance and a closer analogy with NIPALS is proposed. The modification targetsX T X, which is implicitly computed in NIPALS and accumulates the noise inX along its diagonal.