tuning_parameters_CAT.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 (CatBoostRegressor) ---
# used only to modify CatBoostRegressor behavior
# simplified tuning used together with laSFM
'model__learning_rate': [0.03, 0.05], # stable and moderately fast learning
'model__iterations': [300], # fixed; early stopping avoids overfitting
'model__depth': [6], # typical moderate tree complexity (drop 8 to reduce grid size)
'model__l2_leaf_reg': [1, 5], # preserve regularization extremes (drop middle value)
'model__bagging_temperature': [0.0, 0.5], # randomness in sampling, helps generalization
'model__grow_policy': ['SymmetricTree', 'Lossguide'], # balanced tree-growing strategies
'model__verbose': [0], # avoids verbose from the non-sklearn library

# --- Model tuning (CatBoostRegressor) ---
# used only to modify CatBoostRegressor behavior
#'model__learning_rate': [0.01, 0.03, 0.05, 0.1], # from slow to moderately fast learning rates
#'model__iterations': [300, 500, 1000], # wider range to balance training length and performance
#'model__depth': [4, 6, 8, 10], # shallow to deep trees to control complexity
#'model__l2_leaf_reg': [1, 3, 5, 7, 9], # finer granularity for L2 regularization strength
#'model__bagging_temperature': [0.0, 0.3, 0.5, 1.0], # various sampling randomness levels for generalization
#'model__grow_policy': ['SymmetricTree', 'Lossguide', 'Depthwise'], # added third policy for varied tree growth
#'model__border_count': [32, 50, 100], # number of splits for numerical features, affects granularity
#'model__random_strength': [0.0, 1.0, 2.0], # randomness in score calculation for regularization
#'model__verbose': [0], # quiet mode to suppress output
}