A chemometrician's guide to transfer learning

Abstract

Transfer learning (TL), the sub‐discipline of machine learning devoted to learning from different domains, has gained increasing attention over the past decade. With the current contribution, we aim at giving a concise overview on theory, concepts, and applications of TL from a chemometrician's perspective and draw some connections to previous work on calibration model updating/adaptation and calibration transfer. Furthermore, we provide a demonstration of the application of TL in analytical chemistry and discuss the benefits and challenges associate with its use for real‐world problems. We conclude the paper by discussing some open problems and by contemplating on future research directions.

Dual‐constrained and primal‐constrained principal component analysis

“…This assessment via some embedding fitness criterion is likely to be application dependent, though. For example, embedding a Laplacian matrix based upon UMAP projected distances did not work well (enough) in this paper, but using embedded Laplacian matrices for calibration transfer purposes did work well in other studies 15,17,18 …”

“…For example, regression applications using non‐Euclidean embedding matrices have shown success with the analysis of microbiome data involving nuclear magnetic resonance data 4,35 . In the context of calibration transfer and maintenance, non‐Euclidean Laplacian matrices were custom built for pulling together samples from different instruments into a common calibration domain 17 …”

“…In contrast, we investigate the behavior of linear dimension reduction where the nearest‐neighbor topology has been externally supplied to us by a nonlinear manifold embedding (e.g., UMAP). In this sense, this paper shares common topology‐preserving mechanisms that have been explored in chemometrics, in particular domain adaptation and calibration transfer 15–18 …”

Impact of Regressand Stratification in Dataset Shift Caused by Cross-Validation

“…There are different types of dataset shift that have been studied in the specialized literature, such as covariate shift [11] (which affects the distributions of the input variables), conditional shift [12] (which also affects the conditional distributions of the output variable given an input) and posterior shift [13] (which is produced when the conditional distributions of the output given an input vary but the input distributions do not). Among them, a common type of dataset shift, known as target shift or prior probability shift [11,14], occurs in the output variable.…”

A Federated Classification Approach of Waste Lubricant Oils in Geographically Distributed Laboratories

“…In this context, efforts were devoted to the development of analytical solutions that allow to move the knowledge acquired from one site or data source to another where our interest is focused, such as in the activities of scale-up, product transfer, and process monitoring. , A variety of model transfer approaches were developed for handling specifically spectroscopic data. − Techniques such as partial least-squares (PLS) model inversion, Joint-Y PLS, calibration/model transfer, ,, and, more recently, transfer learning ,,, and domain adaptation offer different paths to achieve the aforementioned goals. These methods address partially the aforementioned challenge of connecting multiple sites, but their objective is not to provide a global management platform but instead a way to transfer existing knowledge to a new context of interest.…”