Microsoft LightGBM
pyLightGBM: python binding forFeatures:
- Regression, Classification (binary, multi class)
- Feature importance (
clf.feature_importance()
) - Early stopping (
clf.best_round
) - Works with scikit-learn:
GridSearchCV
,cross_val_score
, etc... - Silent mode (
verbose=False
)
Installation
Install lastest verion of Microsoft LightGBM then install the wrapper:
pip install git+https://github.com/ArdalanM/pyLightGBM.git
Examples
- Regression:
import numpy as np
from sklearn import datasets, metrics, model_selection
from pylightgbm.models import GBMRegressor
# full path to lightgbm executable (on Windows include .exe)
exec = "~/Documents/apps/LightGBM/lightgbm"
X, y = datasets.load_diabetes(return_X_y=True)
clf = GBMRegressor(exec_path=exec,
num_iterations=100, early_stopping_round=10,
num_leaves=10, min_data_in_leaf=10)
x_train, x_test, y_train, y_test = model_selection.train_test_split(X, y, test_size=0.2)
clf.fit(x_train, y_train, test_data=[(x_test, y_test)])
print("Mean Square Error: ", metrics.mean_squared_error(y_test, clf.predict(x_test)))
- Binary Classification:
import numpy as np
from sklearn import datasets, metrics, model_selection
from pylightgbm.models import GBMClassifier
# full path to lightgbm executable (on Windows include .exe)
exec = "~/Documents/apps/LightGBM/lightgbm"
X, Y = datasets.make_classification(n_samples=200, n_features=10)
x_train, x_test, y_train, y_test = model_selection.train_test_split(X, Y, test_size=0.2)
clf = GBMClassifier(exec_path=exec, min_data_in_leaf=1)
clf.fit(x_train, y_train, test_data=[(x_test, y_test)])
y_pred = clf.predict(x_test)
print("Accuracy: ", metrics.accuracy_score(y_test, y_pred))
- Grid Search:
import numpy as np
from sklearn import datasets, metrics, model_selection
from pylightgbm.models import GBMClassifier
# full path to lightgbm executable (on Windows include .exe)
exec = "~/Documents/apps/LightGBM/lightgbm"
X, Y = datasets.make_classification(n_samples=1000, n_features=10)
gbm = GBMClassifier(exec_path=exec,
metric='binary_error', early_stopping_round=10, bagging_freq=10)
param_grid = {'learning_rate': [0.1, 0.04], 'bagging_fraction': [0.5, 0.9]}
scorer = metrics.make_scorer(metrics.accuracy_score, greater_is_better=True)
clf = model_selection.GridSearchCV(gbm, param_grid, scoring=scorer, cv=2)
clf.fit(X, Y)
print("Best score: ", clf.best_score_)
print("Best params: ", clf.best_params_)
Notebooks
- Using pyLightGBM for Kaggle competition (Allstate Claims Severity)
- [Bayesian global optimization with pyLightGBM using data from Kaggle competition (Allstate Claims Severity)](https://github.com/ArdalanM/pyLightGBM/blob/master/notebooks/bayesian optimization_example_kaggle_allstate.ipynb)
Available parameters (default values):
- application="regression"
- num_iterations=10
- learning_rate=0.1
- num_leaves=127
- tree_learner="serial"
- num_threads=1
- min_data_in_leaf=100
- metric='l2'
- is_training_metric=False
- feature_fraction=1.
- feature_fraction_seed=2
- bagging_fraction=1.
- bagging_freq=0
- bagging_seed=3
- metric_freq=1
- early_stopping_round=0
- max_bin=255
- is_unbalance=False
- num_class=1
- boosting_type='gbdt'
- min_sum_hessian_in_leaf=10
- drop_rate=0.01
- drop_seed=4
- max_depth=-1
- lambda_l1=0.
- lambda_l2=0.
- min_gain_to_split=0.
- verbose=True
- model=None