Selection of important samples for calibration model updating based on kernel coefficients

The predictive performance of PLS models often diminishes for new samples outside the application domain and requires updating. The conventional approach involves augmenting the calibration set with new samples to accommodate changes. However, the efficiency of updating is hampered by the large number of samples in the old calibration set. To expedite the updating process, it is crucial to delete some samples, maintaining a balanced class ratio between old and new samples. Surprisingly, there has been a lack of discussion on how to select old samples thus far. This paper introduces a method known as Kernel Coefficient Selection (KCS), designed to obtain coefficients for each sample in the calibration model and subsequently identify crucial old samples. A large coefficient suggests the sample's significance and the need for retention. The rationale behind this method is grounded in the theory of the duality of two types of models (similarity-based and non-similarity-based). Removing samples with small coefficients from the calibration set still preserves the integrity of the major regression model. Comparative experiments were conducted between models that retained partial important samples and those involving all samples, using both simulated and real soybean meal datasets. The experimental results demonstrate that the KCS method mitigates the issue of number imbalance between old and new samples, thereby enhancing the efficiency of model updating.