Abstract
In the field of chemometrics, an important issue in multivariate calibration is model updating. Model updating is the adaption process in which a model obtained for a given set of samples and measurement conditions (primary) is updated to predict the analyte in new samples and measurement conditions (secondary). Primary and secondary conditions can be different due to variations in the geographical situation, instrumentation, or environment. Model updating can be performed using labeled data sets containing samples with reference analyte values for both conditions. A common approach is performed by sample augmenting the larger primary labeled sample set with a small weighted secondary labeled sample set. In this situation, only one updating model is obtained to predict the analyte amount in both primary and secondary conditions. The proposed new approach is similar to this common approach, but instead of one updated model, two models are formed simultaneously. One model is used to only predict samples from the primary conditions and the second model is based on this primary model but modified relative to the weighted augmented secondary samples. This second model is used to predict samples from the secondary conditions. Both model updating methods require multiple tuning parameters (penalties).
Fine Tuning Model Updating for Multivariate Calibration Maintenance
In the field of chemometrics, an important issue in multivariate calibration is model updating. Model updating is the adaption process in which a model obtained for a given set of samples and measurement conditions (primary) is updated to predict the analyte in new samples and measurement conditions (secondary). Primary and secondary conditions can be different due to variations in the geographical situation, instrumentation, or environment. Model updating can be performed using labeled data sets containing samples with reference analyte values for both conditions. A common approach is performed by sample augmenting the larger primary labeled sample set with a small weighted secondary labeled sample set. In this situation, only one updating model is obtained to predict the analyte amount in both primary and secondary conditions. The proposed new approach is similar to this common approach, but instead of one updated model, two models are formed simultaneously. One model is used to only predict samples from the primary conditions and the second model is based on this primary model but modified relative to the weighted augmented secondary samples. This second model is used to predict samples from the secondary conditions. Both model updating methods require multiple tuning parameters (penalties).