Sum of ranking differences (SRD) to ensemble multivariate calibration model merits for tuning parameter selection and...

Area of Science:

• Chemometrics
• Multivariate data analysis
• Statistical modeling

Background:

• Multivariate calibration methods, such as partial least squares (PLS) and ridge regression (RR), require careful selection of tuning parameters.
• Current methods often rely on single metrics like root mean square error of cross-validation (RMSECV), which can be insufficient for optimal parameter selection.
• Balancing bias, variance, selectivity, and sensitivity is crucial for robust model performance.

Purpose of the Study:

• To introduce and evaluate the Sum of Ranking Differences (SRD) as a novel method for selecting optimal tuning parameters in multivariate calibration.
• To demonstrate SRD's ability to ensemble multiple model evaluation merits for improved parameter selection and model comparison.
• To provide a user-guided approach for automatic model selection based on desired bias-variance trade-offs.

Main Methods:

• The Sum of Ranking Differences (SRD) method was developed to combine and rank multiple model evaluation merits.
• SRD was applied to near-infrared spectral and quantitative structure-activity relationship (QSAR) datasets.
• Partial Least Squares (PLS) and Ridge Regression (RR) were used as example calibration methods.

Main Results:

• SRD provides a consensus ranking of tuning parameters, enabling automatic selection of the best model.
• The method facilitates simultaneous comparison of different calibration techniques for a given dataset.
• SRD avoids subjective decisions on weighting and combining different performance metrics.

Conclusions:

• The Sum of Ranking Differences (SRD) offers an effective and objective approach for tuning parameter selection in multivariate calibration.
• SRD allows for the automatic selection of models that best balance predictive performance and model complexity.
• This method enhances the reliability and comparability of multivariate calibration models across different datasets and algorithms.