- Model : a fitted function that predicts Y based on X.
- X / Y : the explanatory factors and the predicted response.
- Test lines : lines excluded from training, reserved for model validation.
- Overfitting : a model that reproduces its training data but fails on new data.
- R² / Adjusted R² : proportion of explained variability, with the second term penalizing the number of terms.
- VIF : variance inflation factor; measures the redundancy of one term relative to the others.
- Confusion Matrix : Cross-tabulation of observed and predicted classes, based on a categorical dependent variable.
- Desirability : translation of a satisfaction rating on a scale from 0 to 1. The overall desirability is the product individual preferences.
- Gaussian process : a model that predicts a value and its uncertainty in every respect.
- Bayesian Optimization : sequential selection of tests to be performed, guided by a Gaussian process.
- Production / Exploration : refine around the known optimum, or explore the unknown.
- Expected Gain (EI) : Expected improvement of a proposed trial compared to the current best result. – Leave-One-Out (LOO) Cross-Validation : Each test is treated as if it had not been performed.
- Q² (predictive R²) : R² calculated based on these leave-one-out predictions. The true measure of predictive power. – Cover 95 % : percentage of tests that fall within the stated prediction range.
- Nugget : proportion of variability attributed to measurement and process noise.
- ARD : correlation lengths by factor; a short length indicates a factor to which the model makes the response react strongly.
- Simplex : the set of mixtures whose proportions add up to 1.
- Ternary diagram : Triangular representation of a three-component mixture.
Appendix: Definitions
Reading time