| Abstract | Background and Objectives: The goal of this research was to examine the relationship between the composition and functionality of pea flour using the following machine learning algorithms: linear regression, partial least squares regression (PLSR), Gaussian process regression (GPR), support vector regression, gradient‐boosted decision trees, and a standard feed‐forward neural network.
Findings: In general, linear models outperformed non‐linear models. PLSR provided best fits for prediction of emulsion stability, oil holding capacity, foam stability and foam capacity; but was less effective for solubility and water holding capacity, which were best described by the GPR model. Variable Importance in Projection scores, calculated for each PLSR model, showed that protein and acid detergent fiber were both highly influential in predicting foaming capacity (1.52 and 1.55), foaming stability (1.30 and 1.54), oil holding capacity (1.64 and 1.50), and water holding capacity (1.56 and 1.53). Protein was also highly important in predicting solubility (1.80), alongside starch (1.60) whereas lipid was highly predictive (2.02) for emulsion stability.
Conclusion: Application of machine learning models was successful in relating compositional features of pea flour to functionality.
Significance: and Novelty Using machine learning to predict the functional behavior of pea will aid both breeders and product developers in ingredient selection. |
|---|
