NEWS

pye 0.1.0 (2026-06-04)

Initial Release: First official release of the pye package, providing a unified toolkit for high-dimensional binary classification, feature selection, and covariate adjustment.

Penalized Youden index Estimator (PYE): Introduced an embedded feature selection method for low- and high-dimensional binary classification ($p \gg n$) that directly maximizes a differentiable, Kernel-Smoothed (KS) version of the Youden Index using a standard normal CDF kernel.
Covariate-adjusted Youden Index (covYI): Implemented an adaptive extension to incorporate covariates, allowing for observation-specific thresholding ($t_i = c_i^\top \gamma$) and automated covariate selection.

Accelerated Proximal Gradient (APG): Implemented two efficient optimization algorithms tailored for non-convex and non-smooth objective functions: mmAPG (modified monotone variant) and mnmAPG (non-monotone variant).
Sparsity-Inducing Penalties: Integrated closed-form proximal operators for a wide range of penalty functions, including $L_{1/2}$ norm, $L_1$ (Lasso), Elastic-Net, SCAD, and MCP.

Model Estimation: Added core routines pye_KS_estimation and covYI_KS_estimation to perform simultaneous feature selection and coefficient estimation.
Unified Benchmarking: Included wrapper functions to estimate and compare established high-dimensional binary decision engines under identical data handling: Penalized Logistic Regression (plr_estimation), Penalized Support Vector Machines (psvm_estimation), and AUC-based methods (AucPR_estimation).

Hyperparameter Selection: Added automated $k$-fold cross-validation routines (pye_KS_compute_cv, plr_compute_cv, psvm_compute_cv, AucPR_compute_cv) to optimize tuning parameters ($\lambda$ and $\tau$) across grid searches.
Data Generation & Validation: Included create_sample_with_covariates to generate synthetic high-dimensional datasets with controlled correlation structures.
Simulation Wrappers: Added pye_simulation_study and model_simulation_study to automate repeated train-test splits for evaluating selection stability and performance metrics under varying sparsity constraints.