Optimization of Honey Powder Formulation, Chemical Quality, and Antioxidant Activity Using Adaptive Feature Selection Based Multitarget Learning Approach

This study focused on predictive system using adaptive feature selection multitarget learning (AFS-MTL) to optimize powdered honey formulation through the selection of the most influential variables, including honey content, filler type and ratio, drying method, temperature, time, and supporting additives. The research employs three machine learning random forest, support vector machine (SVM), and XGBoost to predict optimal formulations and identify key features influencing powdered honey quality. The analysis focused on chemical parameters (moisture content, water activity, Hydroxymethylfurfural (HMF), reducing sugars, diastase enzymes, and antioxidant activity: 1,1-Diphenyl-2-Picrylhydrazyl (DPPH), 2,2-Azinobis-3-Ethylbenzothiazoline-6-Sulphonic Acid (ABTS), total phenols, and total flavonoids). Acacia monoflora honey was pretreated by evaporation and pasteurization to reach 20% moisture, maltodextrin and gum arabic were used as fillers, and anti-caking agents were used as materials. Results from stage I analysis demonstrated that the AFS-MTL framework effectively filters suboptimal formulations while improving prediction efficiency and accuracy compared to conventional trial and error methods. This approach has high potential for improving predictive accuracy, formulation efficiency, and product standardization in powdered honey production.