lightgbm 的简单实践案例

import matplotlib.pylab as plt
import seaborn as sns
import lightgbm as lgb
import pandas as pd
import numpy as np

data_df = pd.read_csv('train.csv')
label = data_df['TARGET']
feature = data_df.drop(['TARGET','ID'],axis=1)

data_test = pd.read_csv('test.csv')
data_test_ID = data_test['ID']
data_test_feature = data_test.drop(['ID'],axis=1)

feature_all = pd.concat([feature,data_test_feature])

feature_all = pd.get_dummies(feature_all,
                         dummy_na=True,
                         columns=None)

feature_train = feature_all.iloc[:len(feature),:]
feature_test = feature_all.iloc[len(feature):]


# 训练模型
def train_model(data_X,data_y):
    
    from sklearn.model_selection import train_test_split
    X_train,x_test,Y_train,y_test = train_test_split(data_X,data_y,test_size=0.2,random_state=3)
    
    # 创建成lgb特征的数据集格式,将使加载更快
    lgb_train =  lgb.Dataset(X_train, label=Y_train)
    lgb_eval = lgb.Dataset(x_test, label=y_test, reference=lgb_train) 
    
    
    parameters = {
                  'task': 'train',
                  'max_depth': 15,
                  'boosting_type': 'gbdt',
                  'num_leaves': 20,        # 叶子节点数
                  'n_estimators': 50,
                  'objective': 'binary',
                  'metric': 'auc',
                  'learning_rate': 0.2,
                  'feature_fraction': 0.7, #小于 1.0, LightGBM 将会在每次迭代中随机选择部分特征.
                  'bagging_fraction': 1,   #类似于 feature_fraction, 但是它将在不进行重采样的情况下随机选择部分数据
                  'bagging_freq': 3,       #bagging 的频率, 0 意味着禁用 bagging. k 意味着每 k 次迭代执行bagging        
                  'lambda_l1': 0.5,
                  'lambda_l2': 0,
                  'cat_smooth': 10,        #用于分类特征,这可以降低噪声在分类特征中的影响, 尤其是对数据很少的类别
                  'is_unbalance': False,   #适合二分类。这里如果设置为True，评估结果降低3个点
                  'verbose': 0
                  }
    
    
    evals_result = {}  #记录训练结果所用
    gbm_model = lgb.train(parameters,
                    lgb_train,
                    valid_sets=[lgb_train,lgb_eval],
                    num_boost_round=50,          #提升迭代的次数
                    early_stopping_rounds=5,
                    evals_result=evals_result,
                    verbose_eval=10
                    )


    prediction = gbm_model.predict(x_test,num_iteration=gbm_model.best_iteration)
    from sklearn.metrics import roc_auc_score
    roc_auc_score = roc_auc_score(y_test, prediction)
    print(roc_auc_score)
    return gbm_model,evals_result


model,evals_result = train_model(feature_train,label)

#运行结果
[5]     training's auc: 0.946343        valid_1's auc: 0.94609
[10]    training's auc: 0.950425        valid_1's auc: 0.948894
[15]    training's auc: 0.954869        valid_1's auc: 0.950978
[20]    training's auc: 0.957274        valid_1's auc: 0.951505
[25]    training's auc: 0.958921        valid_1's auc: 0.95193
[30]    training's auc: 0.960303        valid_1's auc: 0.951958
Early stopping, best iteration is:
[24]    training's auc: 0.958674        valid_1's auc: 0.952064

# 可视化训练结果以及模型下的特征重要性
def lgb_importance():
    
    model,evals_result = train_model(feature_train,label)
    
    ax = lgb.plot_metric(evals_result, metric='auc') #metric的值与之前的params里面的值对应
    plt.title('metric')
    plt.show()
    
    feature_names_pd = pd.DataFrame({'column': feature_train.columns,
                                     'importance': model.feature_importance(),
                                     })
    plt.figure(figsize=(10, 15))
    sns.barplot(x="importance", y="column", data=feature_names_pd.sort_values(by="importance", ascending=False))  #按照importance的进行降排序
    plt.title('LightGBM Features')
    plt.tight_layout()


lgb_importance()

训练集和验证集的auc分数对比

 

可视化出的所有特征的重要性，可以给前面数据预处理做一定参考