接着上一篇，用lightgbm建立baseline，发现模型效果得到了很大优化（模型分提升为0.73)
In[211]:
from sklearn.model_selection import cross_val_score,KFold
In[228]:
import lightgbm as lgb
In[229]:
from lightgbm import LGBMClassifier
In[232]:
from lightgbm import early_stopping
In[237]:
from sklearn.metrics import roc_auc_score
In[215]:
bad_chars=[':','""','\\','']
for feature in application_train.columns:
if any(bad_char in feature for bad_char in bad_chars):
print(f"Feature '{feature}'包含非法字符。")
In[216]:
#去掉特殊字符import pandas as pd
import re
def clean_column_names(df):
"""
清理DataFrame列名，去除特殊字符，使其符合JSON格式要求。
参数:
df (pd.DataFrame): 输入的DataFrame。
返回:
pd.DataFrame: 列名已清理的DataFrame。
"""
定义一个函数，用于替换单个列名中的特殊字符
def replace_chars(col_name):
替换掉所有非字母数字和下划线的字符
return re.sub(r'\W+', '_', col_name)
应用替换函数到所有列名
df.columns = [replace_chars(col) for col in df.columns]
return df
In[217]:
application_train=clean_column_names(application_train)
application_test=clean_column_names(application_test)
In[218]:
bad_chars=[':','""','\\','']
for feature in application_train.columns:
if any(bad_char in feature for bad_char in bad_chars):
print(f"Feature '{feature}'包含非法字符。")
In[238]:
def fit(train=application_train, valid=application_test):
"""
模型训练函数，
参数：train训练集
valid测试集
返回值：
valid_auc:验证集上AUC指标
feature_importances:特征重要性
test_results:测试集结果
"""
test = valid.copy()
x_train = train.drop(['SK_ID_CURR', 'TARGET'], axis=1)
y_train = train['TARGET']
五折交叉验证
folds = KFold(n_splits=5, shuffle=True, random_state=1412)
定义变量保存预测结果
oof_preds = np.zeros(y_train.shape[0])
test_preds = np.zeros(test.shape[0])
提取特征名
feature_names = list(x_train.columns)
空数组用于存储特征重要性值
feature_importance_values = np.zeros(len(feature_names))
实例化模型
lgb = LGBMClassifier(n_estimators=10000, early_stopping_round=200, random_state=24)
for fold_idx, (train_idx, valid_idx) in enumerate(folds.split(x_train)):
X = x_train.iloc[train_idx]
y = y_train.iloc[train_idx]
valid_X = x_train.iloc[valid_idx]
valid_y = y_train.iloc[valid_idx]
定义早停回调函数
callbacks = [early_stopping(stopping_rounds=200)]
拟合模型
lgb.fit(X, y, eval_set=[(X, y), (valid_X, valid_y)], callbacks=callbacks)
记录特征重要性
feature_importance_values += lgb.feature_importances_ / folds.n_splits
在验证集上进行预测
proba = lgb.predict_proba(valid_X, num_iteration=lgb.best_iteration_)
oof_preds[valid_idx] = proba[:, 1] # 选择正类概率
test_preds += lgb.predict_proba(test[feature_names], num_iteration=lgb.best_iteration_)[:, 1]
valid_auc = roc_auc_score(y_train, oof_preds)
feature_importances = pd.DataFrame({'feature': feature_names, 'importance': feature_importance_values})
test['TARGET'] = test_preds
return valid_auc, feature_importances, test[['SK_ID_CURR', 'TARGET']]
In[246]:
valid_auc,feature_importance,submission=fit(application_train[:50000],application_test)
#发现报错了Do not support special JSON characters in feature name，原因是有些列名里有特殊的字符,这是get_dummies时产生的
In[247]:
valid_auc #0.7421786519800682
In[248]:
#看下特征重要性
feature_importance.sort_values(by='importance',ascending=False)
In[250]:
submission.to_csv('baseline_model_lightgbm.csv',index=False)#提交后成绩0.73
In[251]:
application_train.to_csv('original_application_train.csv')#保存下结果
application_test.to_csv('original_application_test.csv')