Binning-based Methods for Model Fitting and Data Splitting Improve Machine Learning Imbalanced Regression

Minimum Transformation Binning improves density estimation for imbalanced regression on genomics data, enabling more effective downstream model fitting and objective data splitting.

• Why it matters: Quantitative genomics modeling is hindered by imbalanced targets, non-independent observations, and high computational cost, leaving limited guidance on which methods work best.
• What they did: The study evaluates binning-based density estimation, proposes Minimum Transformation Binning, and compares objective data splitting scoring functions, finding all-against-one absolute relative difference best.
• The result: Reciprocal density weighting significantly outperforms non-weighted and complement density weighting, and reciprocal weighted sampling per epoch reduces model-fitting resources even when performance does not generally improve.