LightGBM

LightGBM

属于boosting算法中的一种，全称为轻量的梯度提升机(Light Gradient Boosting Machine)，由微软于2017年开源出来的一款SOTA Boosting算法框架。

跟XGBoost一样，LightGBM也是GBDT算法框架的一种工程实现，不过更加快速和高效

安装的时候要注意，如果遇到问题，可以参考下面这个巧妙处理方法

https://blog.csdn.net/sinat_36226553/article/details/106109821  （神奇的安装lightgbm时的解决方法）

使用iris数据集为例的代码如下：

import pandas as pd
import lightgbm as lgb
from sklearn.metrics import mean_squared_error
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.datasets import make_classification

# 导入数据
iris = load_iris()
data = iris.data
target = iris.target
X_train, X_test, y_train, y_test =train_test_split(data, target, test_size=0.2)

# 创建模型
gbm = lgb.LGBMRegressor(objective='regression',
                        num_leaves=31,
                        learning_rate=0.05,
                        n_estimators=20)
# 模型训练
gbm.fit(X_train, y_train,eval_set=[(X_test, y_test)],eval_metric='l1',early_stopping_rounds=5)
# 预测测试集
y_pred = gbm.predict(X_test, num_iteration=gbm.best_iteration_)
# 模型评估
print(mean_squared_error(y_test, y_pred) ** 0.5)
# 查看特征重要性
print(list(gbm.feature_importances_))

 

我有点懵逼，这个例子逻辑回归效果为什么要用均方误差，逻辑回归不是回归模型，而是分类模型

#LightGBM回归模型五折交叉验证训练的代码
import time
import numpy as np
import pandas as pd
import lightgbm as lgb
from sklearn.model_selection import KFold
from sklearn.metrics import mean_squared_error

# 训练特征，使用时label要换为实际标签名称
features = [f for f in df.columns if f not in [label]]

# 自定义模型评估方法
def evalerror(pred, df):
    label = df.get_label().values.copy()
    score = mean_squared_error(label, pred)*0.5
    return ('mse', score, False)

# 指定超参数
params = {
    'learning_rate': 0.01,
    'boosting_type': 'gbdt',
    'objective': 'regression',
    'metric': 'mse',
    'sub_feature': 0.7,
    'num_leaves': 60,
    'colsample_bytree': 0.7,
    'feature_fraction': 0.7,
    'min_data': 100,
    'min_hessian': 1,
    'verbose': -1,
}

t0 = time.time()
train_preds = np.zeros(train.shape[0])

# 五折交叉验证训练
kf = KFold(n_splits=5, shuffle=True, random_state=43)
for i, (train_index, valid_index) in enumerate(kf.split(train)):
    print('train for {} epoch...'.format(i))
    train2 = train.iloc[train_index]
    valid2 = train.iloc[valid_index]
    lgb_train = lgb.Dataset(train2[features], train2['total_cost'], categorical_feature=['hy', 'sex', 'pay_type'])
    lgb_valid = lgb.Dataset(valid2[features], valid2['total_cost'], categorical_feature=['hy', 'sex', 'pay_type'])
    model = lgb.train(params,
                    lgb_train,
                    num_boost_round=3000,
                    valid_sets=lgb_valid,
                    verbose_eval=300,
                    feval=evalerror,
                    early_stopping_rounds=100)
    # 特征重要性排序
    feat_importance = pd.Series(model.feature_importance(), index=features).sort_values(ascending=False)
    train_preds[valid_index] += model.predict(valid2[features], num_iteration=model.best_iteration)

print('Validset score: {}'.format(mean_squared_error(labels, train_preds)*0.5))
print('Cross Validation spend {} seconds'.format(time.time() - t0))

 参数补充：

max_depth, default=-1, type=int，树的最大深度限制，防止过拟合
min_data_in_leaf, default=20, type=int, 叶子节点最小样本数，防止过拟合
feature_fraction, default=1.0, type=double, 0.0 < feature_fraction < 1.0,随机选择特征比例，加速训练及防止过拟合
feature_fraction_seed, default=2, type=int，随机种子数，保证每次能够随机选择样本的一致性
bagging_fraction, default=1.0, type=double, 类似随机森林，每次不重采样选取数据
lambda_l1, default=0, type=double, L1正则
lambda_l2, default=0, type=double, L2正则
min_split_gain, default=0, type=double, 最小切分的信息增益值
top_rate, default=0.2, type=double，大梯度树的保留比例
other_rate, default=0.1, type=int，小梯度树的保留比例
min_data_per_group, default=100, type=int，每个分类组的最小数据量
max_cat_threshold, default=32, type=int，分类特征的最大阈值