Domain---Invariant Partial Least Squares Regression: A Novel Calibration Transfer Paradigm.

Calibration transfer (CT) is an important and widely studied problem in chemometrics. However, most of the current CT methods (e.g. DS, PDS) are limited in terms of application scope, i.e. they learn a mapping between the signals of calibration standards recorded on two devices in order to correct for the difference in the instruments’ response [1]. In the absence of calibration standards, adaptation of calibration models is usually achieved by acquiring additional reference measurements in the target domain in order to account for any source of new variation [2]. In the current contribution, we introduce a novel extension to partial least squares (PLS) regression involving domain regularization in order to enforce that source (master) and target (slave) domain data are aligned in the latent variable subspace (Figure 1). Notably, the proposed extensions allow incorporation of labeled and unlabeled data (i.e. data without reference values) from source and target domains (semi-supervised learning) with the option to adapt source calibration models completely unsupervised, without the need for reference measurements in the target domain. The main focus of this contribution will be set on basic concepts of domain regularization, technical aspects of domain-invariant latent variable extraction and model selection. Application of domain-invariant PLS (di-PLS) will be exemplified on a real-world near infrared spectroscopy dataset from a Melamine Formaldehyde resin production plant, where the aim is to adapt a source calibration model to a target domain with altered MF composition.