Improved independent component regression modeling

The conventional independent component regression (ICR), as an exclusive two-step implementation algorithm, has the risk similar to principal component regression (PCR). That is, the extracted independent components (ICs) are not guaranteed to be informative with respect to quality prediction and interpretation. Moreover, it inherits some inconveniences of conventional ICA. In this paper, first, the drawbacks of original ICR are analyzed. Then a modified ICR (M-ICR) modeling algorithm is developed. To enhance the causal relationship between the extracted ICs and quality variables, a dual-objective optimization solution is constructed in the first-step feature extraction modeling. It simultaneously considers two-fold statistical requirements, the independence and quality-correlation. Moreover, their different roles in calibration modeling can be quantitatively evaluated by flexibly adjusting the sub-optimization objective weights. The practicability and performance of M-ICR are illustrated and discussed in simulation experiment.