cv-score-predict 0.2.0

Cross-validated ensemble prediction with LGBM, XGBoost, and CatBoost — with safe categorical handling, multi-seed averaging, and artifact return.

cv-score-predict

A robust utility for cross-validated ensemble prediction that performs per‑fold early stopping and exposes flexible prediction outputs for advanced stacking, diagnostics, or custom ensembling.
Each fold trains LightGBM, XGBoost, or CatBoost with early stopping on its validation split; the resulting estimators generate out-of-fold (OOF) and test predictions with configurable aggregation. The function supports custom preprocessing pipelines, dynamic per-fold categorical encoding, repeated CV over multiple seeds, and when requested — returns trained models along with their corresponding fold-specific preprocessors.

Designed for kagglers, ML engineers, and data scientists who need reliable, leakage-free CV with minimal boilerplate.

---

✨ Key Features

• Per‑fold early stopping: Each fold trains with early stopping on its validation split and uses the early‑stopped estimator for OOF and test predictions.
• Flexible prediction structures controlled by return_raw_test_preds:
  • OOF predictions (oof_preds_df): Always one column per (model, seed) — predictions from all folds for a given (model, seed) are stitched together into a single complete column.
  • Test predictions (test_preds_df):
    • Default (return_raw_test_preds=False): Averaged across folds → one column per (model, seed) matching OOF structure. Ideal for direct stacking/blending with OOF predictions.
    • Raw mode (return_raw_test_preds=True): Per-fold predictions → one column per (model, seed, fold). Preserves fold-level variance for diagnostics or custom aggregation.
• Multi-model support: Train LightGBM ('lgb'), XGBoost ('xgb'), and CatBoost ('cb') in the same CV loop.
• Safe fold-wise preprocessing: Accepts any scikit-learn–compatible processor with fit_transform/transform. Fitted independently per fold to prevent data leakage.
• Automatic robust categorical handling (always enabled):
  • Detects object/string/categorical columns after the base processor runs,
  • Fits an OrdinalEncoder per fold with explicit unseen-category handling:
    • Unseen categories → encoded as -1
    • Missing values → encoded as -1
    • Training data guaranteed to contain -1 via encoded_missing_value=-1
  • Converts encoded integers to pandas 'category' dtype for native booster support
  • Automatically sets model-specific flags: enable_categorical=True for XGBoost, cat_features=col_names for CatBoost. LightGBM auto-detects categories from dtype.
  • Critical benefit: Satisfies XGBoost's strict validation (all test categories must exist in training) while handling unseen values gracefully.
• Repeated CV over seeds: Accepts a single seed or a list of seeds; CV is repeated for each seed, and all raw predictions are preserved.
• Flexible scoring and thresholding:
  • Custom scoring_dict supported (e.g., accuracy, log loss, RMSE).
  • Defaults: ROC AUC for classification, RMSE for regression.
  • For classification, return probabilities (predict_proba=True) or binary labels (predict_proba=False) using `decision_threshold.
• Artifact return: When return_trained=True, returns a list of tuples (fold_processor, model) — one per model × fold × seed — where fold_processor is the preprocessor fitted on that fold’s training data,
• Transparent, diagnostic-rich logging: With verbose=2 (default), the function prints:
  • Per-fold scores for every model,
  • Stacked (mean of model predictions) score per fold,
  • Per-seed mean scores (by model and stacked),
  • Final cross-seed summary of mean CV performance.
  • → Enables instant diagnosis of model instability, fold bias, or seed sensitivity — no extra code needed.

---

📥 Parameters

Parameter	Type	Default	Description
X	pd.DataFrame	—	Training features.
y	Union[pd.Series, np.ndarray]	—	Target values.
X_test	Optional[pd.DataFrame]	None	Test set for final prediction. If None, no test predictions are returned.
pred_type	str	—	Either 'classification' or 'regression' (required).
processor	Optional[object]	None	Preprocessing pipeline with fit_transform and transform methods. Must return a pd.DataFrame (use set_output(transform='pandas')). If None, features are passed through unchanged.
models	Union[List[str], str]	('lgb', 'xgb', 'cb')	Models to ensemble. Supported: 'lgb' (LightGBM), 'xgb' (XGBoost), 'cb' (CatBoost).
params_dict	Optional[Dict[str, dict]]	None	Model-specific hyperparameters. Keys: model names; values: param dicts.
scoring_dict	Optional[Dict[str, Callable]]	None	Metrics for evaluation. Keys: metric names; values: scoring functions (e.g., roc_auc_score). Defaults: {'roc_auc': roc_auc_score} (classification), {'rmse': rmse_fn} (regression).
decision_threshold	float	0.5	Threshold to convert probabilities to class labels (classification only).
n_splits	int	5	Number of cross-validation folds.
random_state	Union[int, List[int]]	42	Seed(s) for reproducibility. If a list, CV is repeated for each seed and results are averaged.
early_stopping_rounds	int	50	Early stopping rounds for boosting models (if not overridden in params_dict).
verbose	int	2	Logging level: 2 = full per-fold details, 1 = final summary, 0 = silent.
return_trained	bool	False	If True, returns a list of (fold_processor, model) tuples (one per model × fold × seed).
predict_proba	bool	True	For classification: if True, return probabilities; if False, return binary labels (using decision_threshold). Ignored for regression.
return_raw_test_preds	bool	False	Controls test prediction structure: - False (default): Average predictions across folds per (model, seed) → matches OOF structure. - True: Return raw per-fold predictions → one column per (model, seed, fold).

---

🚀 Installation

pip install cv-score-predict

Requirements:

• Python ≥ 3.8
• Dependencies: numpy, pandas, scikit-learn ≥1.4, lightgbm, xgboost, catboost

---

📌 Basic Usage

import pandas as pd
from cv_score_predict import cv_score_predict

# Simulate data
X = pd.DataFrame({
    "num": [1, 2, 3, 4, 5, 6, 7, 8],
    "cat": ["A", "B", "A", "C", "B", "A", "C", "D"]
})
y = [0, 1, 0, 1, 1, 0, 1, 0]
X_test = pd.DataFrame({"num": [9, 10], "cat": ["B", "E"]})  # 'E' is unseen

# Run CV with 2 seeds → get OOF and *averaged* test predictions
oof_preds_df, test_preds_df, _ = cv_score_predict(
    X=X,
    y=y,
    X_test=X_test,
    pred_type="classification",
    models=["lgb", "xgb"],
    random_state=[42, 123],
    n_splits=2,
    verbose=2,
)

# Analyze prediction structures
print("OOF predictions shape:", oof_preds_df.shape)   # (8, 4) → 2 models × 2 seeds
print("Test predictions shape:", test_preds_df.shape) # (2, 4) → 2 models × 2 seeds (averaged across folds)
print(oof_preds_df.columns.tolist())
# ['lgb_seed_42', 'xgb_seed_42', 'lgb_seed_123', 'xgb_seed_123']
print(test_preds_df.columns.tolist())
# ['lgb_seed_42', 'xgb_seed_42', 'lgb_seed_123', 'xgb_seed_123'] ← matches OOF!

# Direct stacking: average OOF and test predictions together
final_oof = oof_preds_df.mean(axis=1)
final_test = test_preds_df.mean(axis=1)

---

🔧 Advanced Usage: Reuse Artifacts for New Data

from sklearn.preprocessing import StandardScaler
from sklearn.compose import make_column_transformer
from sklearn.metrics import roc_auc_score, accuracy_score, log_loss
from cv_score_predict import cv_score_predict

# Define a processor that returns a DataFrame
base_processor = make_column_transformer(
    (StandardScaler(), ["num"]),
    remainder="passthrough"
).set_output(transform='pandas')

scoring_dict = {
    "roc_auc": roc_auc_score,
    "accuracy": accuracy_score,
    "log_loss": log_loss,
}
params_dict = {
    "lgb": {"learning_rate": 0.1, "num_leaves": 100},
    "xgb": {"learning_rate": 0.1, "max_depth": 10},
    "cb": {"learning_rate": 0.1, "depth": 8},
}
# Run CV and return artifacts
oof_preds_df, _, trained_pipelines = cv_score_predict(
    X, y,
    X_test=None,
    pred_type="classification",
    processor=base_processor,
    models=["lgb", "xgb", "cb"],
    params_dict=params_dict,
    scoring_dict=scoring_dict,
    random_state=[42, 123],
    n_splits=5,
    return_trained=True,
)
# Create new data with unseen category and missing value
X_new = pd.DataFrame({"num": [7, 8], "cat": [None, "Z"]})

# Transform and predict using each trained pipeline
all_new_preds = []
for fold_processor, model in trained_pipelines:
    X_new_proc = fold_processor.transform(X_new)  # Handles None/'Z' → -1 automatically
    pred = model.predict_proba(X_new_proc)[:, 1]
    all_new_preds.append(pred)

# Ensemble by averaging
final_new_pred = np.mean(all_new_preds, axis=0)

This gives you a leakage-free stacking pipeline with proper early stopping and categorical handling.

---

📝 Notes

• Categorical handling is always active — detection happens after your base processor runs, so processors that create/modify categoricals (e.g., binning) work correctly.
• Column naming conventions:
  • OOF predictions: {model}_seed_{seed}(always)
  • Test predictions:
    • Averaged mode (return_raw_test_preds=False): {model}_seed_{seed} ← matches OOF
    • Raw mode (return_raw_test_preds=True): {model}_seed_{seed}_fold_{fold}
• Averaging happens before thresholding: Probabilities are averaged across folds first, then thresholded (when predict_proba=False). This preserves probability semantics and avoids averaging binary labels.
• Always use .set_output(transform="pandas") in sklearn pipelines to preserve column names and dtypes.

---

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.