# seed=1000 #随机种子
# #eval_metric= 'auc')
# In[300]:
#模型保存
#bst.save_model('0001.model')
##计算训练和测试的ks
train_pred = bst.predict(dtrain)
train_labels = dtrain.get_label()
test_pred = bst.predict(dtest)
train_labels = dtest.get_label()
train_perf = sc.perf_eva(y_train, train_pred, title = "train")
test_perf = sc.perf_eva(y_test, test_pred, title = "test")
# PSI
# 建立机器学习模型
#train_p = train_pred.copy()
#test_p = test_pred.copy()
print("各样本prediction为", preds)
# In[285]:
train_pred.shape
cm =confusion_matrix(y_train, train_pred_new)
TPR = round(cm[0,0]/(cm[0,0]+cm[0,1]),4)
FPR = round(cm[1,0]/(cm[1,0]+cm[1,1]),4)
ks = round(TPR - FPR,4)
#dpd_rate = round(cm[1,0]/(cm[0,0]+cm[1,0]),4)
acc = round((cm[0,0]+cm[1,1])/cm.sum(),4)
print(i, TPR, FPR, ks, acc)
#测试集
#test_pred_new = np.array(pd.DataFrame(test_pred).iloc[:, 0].apply(lambda x: 0 if x<0.35 else 1))
proba_range = np.arange(0.2,0.9,0.01)
#print(proba_range)
print('阈值', '真正率', '假正率', 'ks', '准确率')
for i in proba_range:
i = round(i,2)
test_pred_new = np.array(pd.DataFrame(test_pred).iloc[:, 0].apply(lambda x: 0 if x<i else 1))
cm =confusion_matrix(y_test, test_pred_new)
TPR = round(cm[0,0]/(cm[0,0]+cm[0,1]),4)
FPR = round(cm[1,0]/(cm[1,0]+cm[1,1]),4)
ks = round(TPR - FPR,4)
#dpd_rate = round(cm[1,0]/(cm[0,0]+cm[1,0]),4)
acc = round((cm[0,0]+cm[1,1])/cm.sum(),4)
dpd = round(cm[1,0]/(cm[0,0]+cm[1,0]),4)
print(i, TPR, FPR, ks, acc, dpd)
train_perf = sc.perf_eva(y_train, train_pred, title='train')
test_perf = sc.perf_eva(y_test, test_pred, title='test')
#保存模型
joblib.dump(lr, 'insight.m')
X_train = train_woe.loc[:, train_woe.columns != 'BAD']
y_test = test_woe.loc[:, 'BAD']
X_test = test_woe.loc[:, train_woe.columns != 'BAD']
# Fit logit model
lr = sm.GLM(y_train, X_train, family=sm.families.Binomial())
fit = lr.fit()
fit.summary()
# Get probabilities
train_pred = fit.predict(X_train)
test_pred = fit.predict(X_test)
# Plot diagnositcs
test_perf = sc.perf_eva(y_test, test_pred, title="test", plot_type=['ks'])
test_perf = sc.perf_eva(y_test, test_pred, title="test", plot_type=['roc'])
class ModelDetails():
def __init__(self, intercept, coefs):
"""Because the scorecardpy package can only take a model class of
LogisticRegression from the scikit-learn package this class is needed
to hold the values from the statsmodels package.
"""
self.intercept_ = [intercept]
self.coef_ = [coefs.tolist()]
model = ModelDetails(fit.params[0], fit.params[1:])
def evaluate_main(self):
writer = pd.ExcelWriter("{}_report.xlsx".format(self.filename))
odds0 = float(self.df_train_woe[self.target_name].value_counts()[1]) / float(self.df_train_woe[self.target_name].value_counts()[0])
b = self.double_score / np.log(2)
a = self.base_score + b * np.log(odds0)
card = sc.scorecard(self.bins_adj, self.model, self.final_features, points0=self.base_score,
odds0=odds0,
pdo=self.double_score)
card_df = pd.DataFrame(columns=["variable", "bin", "points"])
for key, value in card.items():
card_df = pd.concat([card_df, value])
card_df.to_excel(writer, 'card_result')
self.train_pred = self.model.predict_proba(self.df_train_woe[self.final_features])[:, 1]
perf = sc.perf_eva(self.df_train_woe[self.target_name], self.train_pred, title="train")
print("On train-data, the evaluation follows:\nks={}, auc={}, gini={}".format(perf["KS"], perf["AUC"],
perf["Gini"]))
perf["pic"].savefig("{}_train.png".format(self.filename))
_score = sc.scorecard_ply(self.df_train, card, print_step=0)
_score["flag"] = self.df_train_woe[self.target_name]
_score["pred"] = self.train_pred
_rs = self._get_score_table(_score, a, b)
_rs.to_excel(writer, 'train_result')
if self.df_test.any().any():
y_test = self.df_test_woe[self.target_name]
self.test_pred = self.model.predict_proba(self.df_test_woe[self.final_features])[:, 1]
perf = sc.perf_eva(y_test, self.test_pred, title="test")
print("On test-data, the evaluation follows:\nks={}, auc={}, gini={}".format(perf["KS"],
perf["AUC"],
perf["Gini"]))
perf["pic"].savefig("{}_test.png".format(self.filename))
_score = sc.scorecard_ply(self.df_test, card, print_step=0)
_score["flag"] = self.df_test_woe[self.target_name]
_score["pred"] = self.test_pred
_rs = self._get_score_table(_score, a, b)
_rs.to_excel(writer, 'test_result')
if self.df_ott.any().any():
y_ott = self.df_ott_woe[self.target_name]
self.ott_pred = self.model.predict_proba(self.df_ott_woe[self.final_features])[:, 1]
try:
perf = sc.perf_eva(y_ott, self.ott_pred , title="ott")
print("On ott-data, the evaluation follows:\nks={}, auc={}, gini={}".format(perf["KS"],
perf["AUC"],
perf["Gini"]))
perf["pic"].savefig("{}_test.png".format(self.filename))
_score = sc.scorecard_ply(self.df_ott, card, print_step=0)
_score["flag"] = self.df_ott_woe[self.target_name]
_score["pred"] = self.ott_pred
_rs = self._get_score_table(_score,a,b)
_rs.to_excel(writer, 'ott_result')
except:
self.log.info("Cannot caculation the ott data!")
importance = {x: y for x, y in zip(self.final_features, self.model.coef_[0])}
iv_df = self.iv_df[self.iv_df['variable'].isin(self.final_features)]
iv_df["coef"] = iv_df.variable.map(lambda x: importance[x])
iv_df.to_excel(writer, 'feature_importance')
writer.close()
self.log.info("全部环节结束，请查看相关文件！")