微网站开发商,免费做效果图的软件,移动应用开发和软件技术的区别,全国购网站建设数据还是采用这个例子里的数据#xff0c;具体背景也同上。 添模型构建——使用逻辑回归构建模型#xff0c;lightGBM进行特征筛选 lightGBM模型介绍请看这个链接#xff1a;集成学习——Boosting算法#xff1a;Adaboost、GBDT、XGBOOST和lightGBM的简要原理和区别 具体代…数据还是采用这个例子里的数据具体背景也同上。 添模型构建——使用逻辑回归构建模型lightGBM进行特征筛选 lightGBM模型介绍请看这个链接集成学习——Boosting算法Adaboost、GBDT、XGBOOST和lightGBM的简要原理和区别 具体代码如下 导入模块
# 导入模块
import pandas as pd
import numpy as np
import lightgbm as lgb
from sklearn.metrics import roc_auc_score, roc_curve, classification_report
from sklearn import metrics
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt
import seaborn as sns
import math读取数据
df pd.read_csv(Bcard.txt)
print(df.info())
df.head()划分训练集和测试集
# 划分测试集和验证集
train df[df.obs_mth!2018-11-30].reset_index().sort_values(obs_mth, ascendingFalse)
val df[df.obs_mth 2018-11-30].reset_index()
train.head()将训练集的数据进行分组
# 按照时间先后顺序分为5组
train[rank] [i for i in range(train.shape[0])]
train[rank] pd.cut(train[rank], bins5, labels[i for i in range(5)])
train[rank].value_counts()获取特征
ft_lst train.columns
ft_lstft_lst.drop([index,rank,bad_ind,obs_mth,uid])
ft_lst定义模型函数
# 先定义lgb模型函数
def lgb_test(train_X,train_y,test_X,test_y):from multiprocessing import cpu_countlgb_clf lgb.LGBMClassifier(learning_rate0.05,n_estimators100)lgb_clf.fit(train_X, train_y, eval_set[(train_X, train_y), (test_X, test_y)], eval_metricauc, early_stopping_rounds100)lgb.plot_metric(lgb_clf,metricauc)
# print(lgb_clf.n_features_)return lgb_clf, lgb_clf.best_score_[valid_1][auc]进行特征交叉筛选
# 使用lightgbm进行特征交叉筛选feature_lst []
ks_train_lst []
ks_test_lst []# 按照组别对其进行特征筛选
for rk in set(train[rank]):test_df train[train[rank]rk]train_df train[train[rank]!rk]train_X train_df[ft_lst]train_y train_df.bad_indtest_X test_df[ft_lst]test_y test_df.bad_indmodel,auc lgb_test(train_X,train_y,test_X,test_y)feature pd.DataFrame({name:model.booster_.feature_name(),importance:model.feature_importances_}).set_index(name)feature_lst.append(feature)pred_y_train model.predict_proba(train_X)[:,1]pred_y_test model.predict_proba(test_X)[:,1]train_fpr, train_tpr,_ roc_curve(train_y, pred_y_train)test_fpr, test_tpr, _ roc_curve(test_y, pred_y_test)train_ks abs(train_fpr-train_tpr).max()test_ks abs(test_fpr-test_tpr).max()train_auc metrics.auc(train_fpr, train_tpr)test_auc metrics.auc(test_fpr, test_tpr)ks_train_lst.append(train_ks)ks_test_lst.append(test_ks)计算ks
print(train_ks, np.mean(ks_train_lst))
print(test_ks, np.mean(ks_test_lst))进行特征筛选
# 将5组特征值组合取平均值并取大于20的特征
feature_importance pd.concat(feature_lst, axis1).mean(axis1)
lst feature_importance[feature_importance20].index.to_list()
lst使用模型构建评分卡
# 使用lightgbm构建评分卡
X train[lst]
y train.bad_indevl_X val[lst]
evl_y val.bad_ind# 训练集的分类评估
model, auc lgb_test(X,y, evl_X, evl_y)
y_pred model.predict_proba(X)[:,1]
train_fpr, train_tpr,_ roc_curve(y, y_pred)
train_ks abs(train_fpr-train_tpr).max()
train_auc metrics.auc(train_fpr, train_tpr)
print(train_ks,train_ks)# 测试集的分类评估
y_pred model.predict_proba(evl_X)[:,1]
test_fpr,test_tpr,_ roc_curve(evl_y, y_pred)
test_ks abs(test_fpr-test_tpr).max()
test_auc metrics.auc(test_fpr, test_tpr)
print(test_ks,test_ks)绘制roc曲线
# 绘制roc曲线
plt.figure(figsize(16,10))
plt.plot(train_fpr, train_tpr,colorblue, labeltrain lgb auc%0.3f%train_auc)
plt.plot(test_fpr,test_tpr,colororange, labeltest lgb auc%0.3f%test_auc)
plt.plot([0,1],[0,1],--, colorblack)
plt.xlabel(false positive rate)
plt.ylabel(true positive rate)
plt.xlim(0,1)
plt.ylim(0,1)
plt.title(ROC Curve)
plt.legend(loc1)
plt.show()跟评分卡公式对其进行评分并划分等级
# 由于lightgbm没有回归系数所以我们采用原始的评分卡公式
def score(p):score 55050*math.log2((1-p)/p)return scoreval[p] model.predict_proba(evl_X)[:,1]
val[score] val.apply(lambda x:score(x.p), axis1)
print(classification_report(evl_y,model.predict(evl_X), target_names[good,bad]))# 根据评分进行分级
def level(score):level if score 600: level D elif score 640 and score 600 : level C elif score 680 and score 640: level B elif score 680 : level A return levelval[level] val.apply(lambda x:level(x.score), axis1)
val.level.value_counts()验证ks
# 验证ks
fpr,tpr,_ roc_curve(evl_y, val[score])
ks abs(fpr-tpr).max()
print(ks)0.42314255674351975 # 生成报告
temp pd.DataFrame()
temp[bad_rate_pred] val[p]
temp[real_bad] evl_y
temp.sort_values(bad_rate_pred, inplaceTrue, ascendingFalse)
temp[num] [i for i in range(temp.shape[0])]
temp[num] pd.cut(temp.num, bins20, labels[i for i in range(20)])report pd.DataFrame()
report[bad] temp.groupby(num).real_bad.sum()
report[good] temp.groupby(num).real_bad.count()-report[bad]
report[bad_cnt] report[bad].cumsum()
report[good_cnt] report[good].cumsum()
good_total report[good_cnt].max()
bad_total report[bad_cnt].max()
report[bad_pct] round(report[bad_cnt]/bad_total,3)
report[good_pct] round(report[good_cnt]/good_total,3)
report[bad_rate] report.apply(lambda x:round(x.bad/(x.goodx.bad), 3), axis1)
def cal_ks(x):ks x.bad_pct - x.good_pctreturn round(math.fabs(ks),3)report[ks] report.apply(cal_ks, axis1)
report绘制bad_rate和ks折线图
# 绘制bad_rate和KS的折线图
fig plt.figure(figsize(16,10))
ax fig.add_subplot(111)
ax.plot(range(20), report[bad_rate],-o,labelbad_rate)
ax2 ax.twinx()
ax2.plot(range(20), report[ks],--o,colorr,labelks)
ax.grid()
ax.set_xlim(-1,20)
ax.set_ylim(0,0.14)
ax2.set_ylim(0,0.5)
ax.set_ylabel(bad_rate)
ax2.set_ylabel(ks)
ax.set_xlabel(num)
ax.legend(loc2)
ax2.legend(loc0)
