Adaptive sample selection for individual test-sample prediction under distribution shift via minimum regularized...

Distribution shift poses a significant challenge for data analysis. If a portion of the training samples does not match the data distribution in the real application, the predictive performance and generalization ability of the model will be reduced. This phenomenon is particularly evident in partial least squares (PLS), a widely employed regression technique for addressing high-dimensional regression problems. To address this challenge, this study introduces the adaptive minimum regularized covariance determinant (AMRCD) method, which adaptively selects training samples from the same distribution as a single test sample. The method facilitates the construction of accurate PLS models and effectively improves prediction accuracy for test samples. Additionally, the AMRCD framework incorporates regularization techniques to guarantee that the covariance matrix remains well-conditioned in high-dimensional spaces. The proposed method exhibits significantly enhanced predictive accuracy compared to classical PLS and an alternative sample selection framework, as validated on three simulations and two real-world datasets.