A Hybrid Machine Learning Framework to Improve Morphological Trait Recovery in Avian Datasets.

Area of Science:

• Ecology
• Evolutionary Biology
• Machine Learning
• Biodiversity Science

Background:

• Missing data in morphological trait datasets hinder ecological and evolutionary research.
• Inaccurate imputation compromises model inference and predictive accuracy.

Purpose of the Study:

• To develop and validate THORBFNN, a novel hybrid imputation framework for accurately recovering missing avian morphological traits.
• To assess THORBFNN's performance against existing imputation methods.

Main Methods:

• A three-stage hybrid imputation framework: regularized K-means clustering, Radial Basis Function Neural Networks (RBFNNs), and hierarchical Bayesian optimization.
• Regularized K-means partitions species into clusters to preserve local morphological structure.
• RBFNNs model nonlinear trait dependencies, with hyperparameters optimized via Bayesian optimization.

Main Results:

• THORBFNN outperformed K-nearest neighbors and Random Forest imputation on a global avian trait dataset (10,000+ individuals, 11 traits).
• Achieved higher R-squared (0.9003) and lower errors (RMSE=0.1652, MAE=0.1096) for four focal traits.
• Ablation studies confirmed THORBFNN captures genuine trait covariation, not statistical artifacts.

Conclusions:

• THORBFNN provides an accurate and efficient method for imputing missing avian morphological data.
• The framework requires no phylogenetic information and scales to large datasets.
• Offers a practical approach for integrating machine learning into biodiversity trait analysis.