银行风控案例-python学习笔记

风控模型开发流程

·        数据抽取

·        数据探索

·        建模数据准备

·        变量选择

·        模型开发与验证

·        模型部署

·        模型监督

加载包：

importos

importsys

importstring

importpymysql

importnumpyasnp

importpandasaspd

importstatsmodels.apiassm

importmatplotlib.pyplotasplt

%matplotlibinline

importseabornassns

importsklearn.cross_validationascross_validation

importsklearn.treeastree

importsklearn.ensembleasensemble

importsklearn.linear_modelaslinear_model

importsklearn.svmassvm

importsklearn.feature_selectionasfeature_selection

importsklearn.metricsasmetrics

 

1. 数据抽取

model_data = pd.read_csv("credit_develop.csv")

model_data.head()#查看数据格式

读取数据后查看大致数据的情况，主要看每个字段的格式属性。比如Branch_of_Bank等字段为分类变量，后续需要进行虚拟化变量处理。

1. 数据探索

数据探索是建模人员了解特征时使用的方法，可以通过数据表或是图形的方式了解整体数据。

model_data.describe().T

查看数据的分布。

data = model_data["Credit_Score"].dropna() # 去除缺失值

sns.distplot(data)

查看信用分数的分布情况

plt.boxplot(data)

利用箱线图查看数据离散情况

model_data=model_data.drop_duplicates()#去除重复项

填充缺失值

model_data = model_data.fillna(model_data.mean()) #用均值来填充

变量相似度分析，变量聚类

simpler = np.random.randint(0,len(model_data),size=50)

sns.clustermap(model_data.iloc[simpler,3:].T,col_cluster=False,row_cluster=True)

1. 生成模型训练/测试数据集

将分类变量变为虚拟变量

Area_Classification_dummy = pd.get_dummies(model_data["Area_Classification"],prefix="Area_Class")

model_data.join(Area_Classification_dummy)

model_data.join(model_data[="Branch"))

·        分成目标变量和应变量

target = model_data["target"]

pd.crosstab(target,"target")

data = model_data.ix[ :,'Age':]

·        分成训练集和测试集，比例为6:4

train_data, test_data, train_target, test_target = cross_validation.train_test_split(data, target, test_size=0.4, random_state=0)

1. 筛选变量

因为使用的是普通的罗吉斯回归，所以变量筛选变得尤为重要。如果筛选不当会产生过拟合或是欠拟合现象。（当然可以使用一些更高级的算法完成筛选功能）首先使用最原始的方法线性相关系数。

corr_matrix = model_data.corr(method='pearson')

corr_matrix = corr_matrix.abs()

sns.set(rc={"figure.figsize": (10, 10)})

sns.heatmap(corr_matrix,square=True,cmap="Blues")

corr = model_data.corr(method='pearson').ix["target"].abs()

corr.sort(ascending=False)

corr.plot(kind="bar",title="corr",figsize=[12,6])

·        使用随机森林方法来选择模型模型变量

rfc = ensemble.RandomForestClassifier(criterion='entropy', n_estimators=3, max_features=0.5, min_samples_split=5)

rfc_model = rfc.fit(train_data, train_target)

rfc_model.feature_importances_

rfc_fi = pd.DataFrame()

rfc_fi["features"] = list(data.columns)

rfc_fi["importance"] = list(rfc_model.feature_importances_)

rfc_fi=rfc_fi.set_index("features",drop=True)

rfc_fi.sort_index(by="importance",ascending=False).plot(kind="bar",title="corr",figsize=[12,6])

5.模型训练

• 使用原始变量进行logistic回归

In [39]: 

logistic_model = linear_model.LogisticRegression()

logistic_model.fit(train_data, train_target)

Out[39]:

LogisticRegression(C=1.0, class_weight=None, dual=False, fit_intercept=True,          intercept_scaling=1, penalty='l2', random_state=None, tol=0.0001)

In [40]: 

test_est = logistic_model.predict(test_data)

train_est = logistic_model.predict(train_data)

In [41]: 

test_est_p = logistic_model.predict_proba(test_data)[:,1]

train_est_p = logistic_model.predict_proba(train_data)[:,1]

In [42]: 

print metrics.classification_report(test_target, test_est)

             precision    recall  f1-score   support          0       0.68      0.61      0.64      2825          1       0.64      0.71      0.67      2775avg / total       0.66      0.66      0.66      5600

In [43]: 

print metrics.classification_report(train_target, train_est)

             precision    recall  f1-score   support          0       0.67      0.63      0.65      4175          1       0.66      0.70      0.68      4225avg / total       0.67      0.67      0.66      8400

In [44]: 

metrics.zero_one_loss(test_target, test_est)

Out[44]:

0.34053571428571427

In [45]: 

metrics.zero_one_loss(train_target, train_est)

Out[45]:

0.33476190476190482

• 目标样本和非目标样本的分数分布

In [46]:

x

red, blue = sns.color_palette("Set1",2)

In [47]: 

sns.kdeplot(test_est_p[test_target==1], shade=True, color=red)

sns.kdeplot(test_est_p[test_target==0], shade=True, color=blue)

Out[47]:

<matplotlib.axes._subplots.AxesSubplot at 0x2183aa20>

• ROC曲线

In [48]: 

fpr_test, tpr_test, th_test = metrics.roc_curve(test_target, test_est_p)

fpr_train, tpr_train, th_train = metrics.roc_curve(train_target, train_est_p)

plt.figure(figsize=[6,6])

plt.plot(fpr_test, tpr_test, color=blue)

plt.plot(fpr_train, tpr_train, color=red)

plt.title('ROC curve')

Out[48]:

<matplotlib.text.Text at 0x2201a940>

In [49]: 

test_x_axis = np.arange(len(fpr_test))/float(len(fpr_test))

train_x_axis = np.arange(len(fpr_train))/float(len(fpr_train))

plt.figure(figsize=[6,6])

plt.plot(fpr_test, test_x_axis, color=blue)

plt.plot(tpr_test, test_x_axis, color=blue)

plt.plot(fpr_train, train_x_axis, color=red)

plt.plot(tpr_train, train_x_axis, color=red)

plt.title('KS curve')

Out[49]:

<matplotlib.text.Text at 0x22171198>

1. 模型的监督

·        分数分布稳定性

分数指分数越高的人违约概率越低，分数分布指的是在分数较低的人群中违约概率是否成正比的。

·        变量判别能力稳定性

指变量的重要性

·        变量分布稳定性

2. 其他机器学习算法

lr = linear_model.LogisticRegression()

lr_scores = cross_validation.cross_val_score(lr, train_data, train_target, cv=5)

print("logistic regression accuracy:")

print(lr_scores)

​

clf = tree.DecisionTreeClassifier(criterion='entropy', max_depth=8, min_samples_split=5)

clf_scores = cross_validation.cross_val_score(clf, train_data, train_target, cv=5)

print("decision tree accuracy:")

print(clf_scores)

​

rfc = ensemble.RandomForestClassifier(criterion='entropy', n_estimators=3, max_features=0.5, min_samples_split=5)

rfc_scores = cross_validation.cross_val_score(rfc, train_data, train_target, cv=5)

print("random forest accuracy:")

print(rfc_scores)

​

etc = ensemble.ExtraTreesClassifier(criterion='entropy', n_estimators=3, max_features=0.6, min_samples_split=5)

etc_scores = cross_validation.cross_val_score(etc, train_data, train_target, cv=5)

print("extra trees accuracy:")

print(etc_scores)

​

gbc = ensemble.GradientBoostingClassifier()

gbc_scores = cross_validation.cross_val_score(gbc, train_data, train_target, cv=5)

print("gradient boosting accuracy:")

print(gbc_scores)

​

svc = svm.SVC()

svc_scores = cross_validation.cross_val_score(svc, train_data, train_target, cv=5)

print("svm classifier accuracy:")

print(svc_scores)

​

abc = ensemble. AdaBoostClassifier(n_estimators=100)

abc_scores = cross_validation.cross_val_score(abc, train_data, train_target, cv=5)

print("abc classifier accuracy:")

print(abc_scores)

logistic regression accuracy:

[ 0.66785714  0.65416667  0.65357143  0.65535714  0.65833333]

decision tree accuracy:

[ 0.74107143  0.75654762  0.73988095  0.73035714  0.73690476]

random forest accuracy:

[ 0.7125      0.72142857  0.71190476  0.72083333  0.68809524]

extra trees accuracy:

[ 0.66964286  0.70714286  0.6827381   0.67738095  0.67797619]

gradient boosting accuracy:

[ 0.78214286  0.78035714  0.77202381  0.76071429  0.75119048]

svm classifier accuracy:

[ 0.50297619  0.50595238  0.50654762  0.50297619  0.50297619]

abc classifier accuracy:

[ 0.76785714  0.76488095  0.76904762  0.74940476  0.75      ]