A hybrid feature selection-based machine learning framework for robust prediction of parthenocarpy in Hibiscus sabdariffa L

Modeling the phenotypic expression of parthenocarpy in Hibiscus sabdariffa is important for improving fruit set, enhancing yield stability, and supporting breeding research. This study presents a hybrid feature selection and machine learning framework that integrates filter-based and regularization-based techniques to identify biologically informative phenotypic traits associated with parthenocarpy, improve model interpretability, and reduce feature dimensionality. Regression models, including Support Vector Machine with a radial basis function kernel (SVM-RBF), Decision Tree, Random Forest, Ensemble Learning and Linear Regression, were developed and evaluated using the complete set of 24 experimentally measured morphological, fruit-, capsule-, and seed-related traits, as well as the compact feature subsets identified by the proposed hybrid feature selection framework. Model performance was assessed using the coefficient of determination (R 2 ), root mean square error (RMSE), mean absolute error (MAE) and concordance correlation coefficient (CCC). The results show that high phenotype-based modeling performance can be achieved using fewer, biophysically relevant traits; however, non-linear models consistently outperformed linear models across all evaluated models (test R 2 values up to 0.983, CCC > 0.99). Among the traits most strongly associated with the final expression of parthenocarpy were seed-related characteristics (i.e., the number of mature and immature seeds, seed weight and seed size), while fruit-related traits (e.g., fruit and capsule volume) provided secondary information. The proposed framework enables dimensionality reduction, improves model interpretability, and provides an interpretable framework for phenotype-based trait prioritization and modeling of complex reproductive traits in Hibiscus sabdariffa . Rather than serving as an early prediction tool prior to seed development, the framework is intended to facilitate phenotype-based analysis and prioritization of traits associated with the observed expression of parthenocarpy.