tuning_parameters_EN.txt

{
# when combined with any model or pipeline, minimizing the number of hyperparameter combinations helps reduce overall search time

# --- Feature selection (SelectKBest) ---
# used only to modify selected SKB behavior
#'feature_selection__k': [25, 50], # number of top features to keep
#'feature_selection__score_func': [mutual_info_regression], # score function (avoid chi2 in a context of regression and f_regression it captures only linear dependancies)

# --- Feature selection (SelectFromModel with lasso) ---
# used only to modify laSFM behavior
#'feature_selection__threshold': [-float('inf')], # disable thresholding to rely solely on max_features
#'feature_selection__max_features': [25, 50], # select exactly this number of top features
#'feature_selection__estimator__alpha': [0.01, 0.1, 1.0], # increase regularization to speed up convergence
#'feature_selection__estimator__max_iter': [500], # reduce iterations for faster training
#'feature_selection__estimator__tol': [1e-2], # relax convergence criteria to save time
#'feature_selection__estimator__fit_intercept': [True], # whether to estimate the intercept

# --- Feature selection (SelectFromModel with elasticnet) ---
# used only to modify enSFM behavior
#'feature_selection__threshold': [-float('inf')], # rank features by importance
#'feature_selection__max_features': [25, 50], # number of top features to keep
#'feature_selection__estimator__alpha': [0.1], # moderate regularization strength
#'feature_selection__estimator__l1_ratio': [0.5], # balanced L1/L2 penalty
#'feature_selection__estimator__max_iter': [300], # max iterations for convergence
#'feature_selection__estimator__tol': [1e-2], # relaxed convergence tolerance

# --- Feature selection (SelectFromModel with ridge) ---
# used only to modify riSFM behavior
'feature_selection__threshold': [-float('inf')], # disable hard thresholding; rank by |coef|
'feature_selection__max_features': [25, 50], # keep top-ranked features only
'feature_selection__estimator__alpha': [0.1, 1.0, 10.0], # regularization strength
'feature_selection__estimator__max_iter': [1000], # ensure convergence for high-dimensional OHE
'feature_selection__estimator__tol': [1e-3, 1e-4], # trade-off speed vs numerical precision
'feature_selection__estimator__fit_intercept': [True], # center target; recommended for regression

# --- Feature selection (SelectFromModel with random forest) ---
# used only to modify rfSFM behavior
#'feature_selection__threshold': [-float('inf')], # rank all features by importance
#'feature_selection__max_features': [25, 50], # number of top features to keep
#'feature_selection__estimator__n_estimators': [100], # number of trees
#'feature_selection__estimator__max_depth': [10], # shallow trees for speed

# --- Model tuning (ElasticNet) ---
# used only to modify ElasticNet behavior
# simplified tuning used together with rfSFM
'model__alpha': [0.001, 0.01], # light to moderate regularization
'model__l1_ratio': [0.2, 0.5, 0.8], # from mostly L2 to balanced to mostly L1
'model__max_iter': [5000], # sufficient for convergence under typical conditions
'model__tol': [1e-4], # moderate tolerance for faster convergence

# --- Model tuning (ElasticNet) ---
# used only to modify ElasticNet behavior
#'model__alpha': [0.0001, 0.001, 0.01, 0.1, 1.0], # extended with 1.0 for stronger regularization
#'model__l1_ratio': [0.05, 0.1, 0.2, 0.3, 0.5, 0.7, 0.9], # finer balance between L1 and L2 regularization
#'model__max_iter': [10000, 15000, 20000], # enough iterations for convergence on larger data
#'model__tol': [1e-6, 1e-5, 1e-4, 1e-3], # lower to moderate tolerances for convergence flexibility
}