shap-hypetune

0.2.7 verified Fri May 01 auth: no python

A Python package for simultaneous hyperparameters tuning and feature selection for gradient boosting models (LightGBM, XGBoost, CatBoost, Random Forest). Current version: 0.2.7. Release cadence: occasional, last release Jun 2023.

pip install shap-hypetune

Common errors

error ModuleNotFoundError: No module named 'shap_hypetune' ↓

cause Library not installed or wrong Python environment.

fix

pip install shap-hypetune

error AttributeError: 'BoostingRFA' object has no attribute 'get_support' ↓

cause Using version older than 0.2.6 where get_support() was added.

fix

Upgrade to shap-hypetune>=0.2.6: pip install --upgrade shap-hypetune

error ValueError: The estimator should be a string among 'lgbm', 'xgb', 'catboost', 'rf' ↓

cause Passed an estimator string that is not recognized or a custom object incorrectly formatted.

fix

Use one of the supported strings: 'lgbm', 'xgb', 'catboost', 'rf' (for Random Forest).

error ImportError: cannot import name 'RFA' from 'shap_hypetune' ↓

cause Trying to import the old alias 'RFA' which was renamed to 'BoostingRFA'.

fix

Use: from shap_hypetune import BoostingRFA

Warnings

breaking In v0.2.7, numpy types np.bool, np.int, np.float were replaced by native Python types. If you rely on exact numpy dtypes in model outputs, upgrade numpy to >=1.20 and check your custom callbacks. ↓

fix Upgrade to v0.2.7 or cast numpy types explicitly.

gotcha The 'n_iter' parameter controls both tuning and feature selection iterations; setting it too low may result in poor convergence. In BoostingRFA and BoostingRFE, n_iter is the total number of trials (if hyperopt=True) or grid points (if grid=True). ↓

fix Use at least 10-20 iterations for tuning; increase if using hyperopt=True.

gotcha Feature selection methods (RFA, RFE) require a fitted model with feature importance; if you pass a custom estimator that does not provide 'feature_importances_', it will raise an error. ↓

fix Use only LightGBM, XGBoost, CatBoost, or Random Forest as the base estimator, or ensure custom estimator has the attribute.

deprecated Usage of 'estimator' parameter as a string (e.g., 'lgbm') is stable, but passing a scikit-learn estimator object is experimental and may break in future versions. ↓

fix Stick to string names for now; if you need a custom estimator, test thoroughly.

Imports

BoostingRFA
wrong
```
from shap_hypetune import RFA
```
correct
```
from shap_hypetune import BoostingRFA
```
RFA was added in v0.2.4 but the class is actually named BoostingRFA; importing just 'RFA' fails

BoostingRFE

wrong

from shap_hypetune.rfe import BoostingRFE

correct

from shap_hypetune import BoostingRFE

Top-level import works; no submodule needed

BoostingRFECV
```
from shap_hypetune import BoostingRFECV
```

BoostingSelector

wrong

from shap_hypetune.selector import BoostingSelector

correct

from shap_hypetune import BoostingSelector

Top-level import is correct; submodule not directly exposed

Quickstart

Basic usage: perform recursive feature addition with hyperparameter tuning using LightGBM. Note: estimator can be 'lgbm', 'xgb', 'catboost', or 'rf' (Random Forest, added in v0.2.4).

import pandas as pd
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from shap_hypetune import BoostingRFA

X, y = make_classification(n_samples=100, n_features=20, random_state=0)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)

model = BoostingRFA(estimator='lgbm', n_iter=5, early_stopping_rounds=10, random_state=0)
model.fit(X_train, y_train)
print('Train accuracy:', model.score(X_train, y_train))
print('Selected features:', model.get_support())