Unsupervised model adaptation for multivariate calibration by domain adaptation-regularization based kernel partial...

Area of Science:

• Chemometrics
• Machine Learning
• Data Science

Background:

• Calibration model adaptation is crucial when training and test data distributions differ.
• Domain-invariant feature representation is a key strategy for effective model adaptation.
• Existing methods may not fully address distribution discrepancies in complex datasets.

Purpose of the Study:

• To propose a novel nonlinear unsupervised model adaptation method, DarKPLS.
• To minimize distribution differences between source and target data in a latent space.
• To enhance the performance of chemometric calibration models through domain alignment.

Main Methods:

• Developed Domain Adaptation Regularization-based Kernel Partial Least Squares Regression (DarKPLS).
• Utilized reproducing kernel Hilbert space (RKHS) for feature projection.
• Aligned distributional means and variances between source and target latent variables.
• Integrated Projected Maximum Mean Discrepancy (PMMD) for fine-grained domain alignment.

Main Results:

• DarKPLS demonstrated improved prediction accuracy on γ-polyglutamic acid fermentation, tobacco, and corn datasets.
• The method showed superior performance compared to Partial Least Squares (PLS), Null Augmented Regression (NAR), Extended Joint Trained (ExtJT), Scatter Component Analysis (SCA), and Domain-Invariant Iterative Partial Least Squares (DIPALS).
• Fine-grained domain alignment via PMMD significantly enhanced adaptation performance.

Conclusions:

• DarKPLS offers an effective nonlinear unsupervised approach for calibration model adaptation in chemometrics.
• The method successfully addresses distribution shifts by aligning latent variable distributions.
• DarKPLS provides a robust solution for improving predictive model performance across different data domains.