20200317_決策樹預測貸款申請

使用決策樹，預測貸款申請

import pandas as pd
#  忽略彈出的warnings
import warnings
warnings.filterwarnings('ignore')  

text=pd.read_excel('data/LoanStats_securev1_2019Q4.xlsx')
text.head()

	id	loan_amnt	funded_amnt	funded_amnt_inv	term	int_rate	installment	grade	sub_grade	emp_title	...	num_tl_90g_dpd_24m	num_tl_op_past_12m	pct_tl_nvr_dlq	percent_bc_gt_75	pub_rec_bankruptcies	tax_liens	tot_hi_cred_lim	total_bal_ex_mort	total_bc_limit	total_il_high_credit_limit
0	164027473	20000	20000	20000	36 months	0.1240	668.12	B	B4	NaN	...	0	2	100.0	50.0	1	0	60800	42566	5200	40000.0
1	163984413	16500	16500	16500	60 months	0.1033	353.27	B	B1	NaN	...	0	0	100.0	0.0	0	0	223390	40913	40500	39890.0
2	164193225	7500	7500	7500	36 months	0.1240	250.55	B	B4	Rn	...	0	7	54.5	16.7	0	0	138468	102122	47700	90768.0
3	162948736	19000	19000	18975	36 months	0.0646	581.99	A	A1	Tech Ops Analyst	...	0	0	100.0	40.0	0	0	184034	28461	38400	35000.0
4	164161686	10000	10000	10000	36 months	0.2055	374.45	D	D2	Planner	...	0	2	100.0	16.7	0	0	639373	161516	24600	172818.0

5 rows × 114 columns

目標變量

text['loan_status'].value_counts()

Current               122625
Fully Paid              3539
In Grace Period         1079
Late (31-120 days)       509
Late (16-30 days)        304
Charged Off               80
n                          1
Name: loan_status, dtype: int64

#0爲已經完成的
def function(x):
    if 'Current' in x:
        return 0
    elif 'Fully Paid' in x:
        return 0
    else:
        return 1
text['loan_status']=text.apply(lambda x:function(x['loan_status']),axis=1)

text['loan_status'].value_counts()

0    126164
1      1973
Name: loan_status, dtype: int64

pos_trainDf = text[text['loan_status'] == 1]
neg_trainDf = text[text['loan_status'] == 0].sample(n=4000, random_state=2018)
text = pd.concat([pos_trainDf, neg_trainDf], axis=0).sample(frac=1.0,random_state=2018)

text.info()

<class 'pandas.core.frame.DataFrame'>
Int64Index: 5973 entries, 18821 to 92872
Columns: 114 entries, id to total_il_high_credit_limit
dtypes: datetime64[ns](1), float64(36), int64(50), object(27)
memory usage: 5.2+ MB

缺失值查看

check_null = text.isnull().sum(axis=0).sort_values(ascending=False)/float(len(text)) #查看缺失值比例
print(check_null[check_null >0.2]) # 查看缺失比例大於20%的屬性。

desc                              0.999833
mths_since_last_record            0.899046
verification_status_joint         0.880629
annual_inc_joint                  0.864055
dti_joint                         0.864055
mths_since_recent_bc_dlq          0.794408
mths_since_last_major_derog       0.769965
mths_since_recent_revol_delinq    0.703164
mths_since_last_delinq            0.548468
dtype: float64

thresh_count = len(text)*0.4 # 設定閥值
data = text.dropna(thresh=thresh_count, axis=1 ) #若某一列數據缺失的數量超過閥值就會被刪除

data.info()

<class 'pandas.core.frame.DataFrame'>
Int64Index: 5973 entries, 18821 to 92872
Columns: 106 entries, id to total_il_high_credit_limit
dtypes: datetime64[ns](1), float64(30), int64(50), object(25)
memory usage: 4.9+ MB

刪除無意義的列

sub_grade：與Grade的信息重複

emp_title ：缺失值較多，同時不能反映借款人收入或資產的真實情況

zip_code：地址郵編，郵編顯示不全，沒有意義

addr_state：申請地址所屬州，不能反映借款人的償債能力

last_credit_pull_d ：LendingClub平臺最近一個提供貸款的時間，沒有意義

policy_code ： 變量信息全爲1

pymnt_plan 基本是n

title： title與purpose的信息重複，同時title的分類信息更加離散

next_pymnt_d : 下一個付款時間，沒有意義

policy_code : 沒有意義

collection_recovery_fee: 全爲0，沒有意義

earliest_cr_line : 記錄的是借款人發生第一筆借款的時間

issue_d ： 貸款發行時間，這裏提前向模型泄露了信息

last_pymnt_d、collection_recovery_fee、last_pymnt_amnt： 預測貸款違約模型是貸款前的風險控制手段，這些貸後信息都會影響我們訓練模型的效果，在此將這些信息刪除

drop_list = ['sub_grade', 'emp_title',  'title', 'zip_code', 'addr_state', 
             'mths_since_last_delinq' ,'initial_list_status','title','issue_d','last_pymnt_d','last_pymnt_amnt',
             'next_pymnt_d','last_credit_pull_d','policy_code','collection_recovery_fee', 'earliest_cr_line']
data.drop(drop_list, axis=1, inplace = True)

data.head()

	id	loan_amnt	funded_amnt	funded_amnt_inv	term	int_rate	installment	grade	emp_length	home_ownership	...	num_tl_90g_dpd_24m	num_tl_op_past_12m	pct_tl_nvr_dlq	percent_bc_gt_75	pub_rec_bankruptcies	tax_liens	tot_hi_cred_lim	total_bal_ex_mort	total_bc_limit	total_il_high_credit_limit
18821	163425898	4500	4500	4500	36 months	0.1612	158.48	C	NaN	RENT	...	0	2	100.0	28.6	0	0	44700	10872	32800	0.0
61234	161908366	20000	20000	20000	60 months	0.2305	564.39	D	NaN	OWN	...	0	0	100.0	33.3	0	0	54349	19572	10400	22349.0
119781	159901427	10000	10000	10000	60 months	0.1862	257.32	D	6 years	OWN	...	0	3	100.0	0.0	0	0	69077	48184	9600	49477.0
49201	162292591	21000	21000	21000	60 months	0.1430	491.91	C	< 1 year	RENT	...	0	0	100.0	0.0	0	0	109894	66662	33800	67194.0
53727	162154208	40000	40000	40000	60 months	0.0819	814.70	A	10+ years	RENT	...	0	0	100.0	50.0	0	0	207370	160985	98000	61725.0

5 rows × 91 columns

分類變量

objectColumns = data.select_dtypes(include=["object"]).columns
data[objectColumns].isnull().sum().sort_values(ascending=False)

emp_length             572
application_type         1
url                      1
total_acc                0
delinq_2yrs              0
purpose                  0
pymnt_plan               0
verification_status      0
annual_inc               0
home_ownership           0
grade                    0
term                     0
dtype: int64

# data['int_rate'] = data['int_rate'].str.rstrip('%').astype('float')
# data['revol_util'] = data['revol_util'].str.rstrip('%').astype('float')
# data['annual_inc'] = data['annual_inc'].str.replace(",","").astype('float')

import numpy as np
objectColumns = data.select_dtypes(include=["object"]).columns # 篩選數據類型爲object的數據
data[objectColumns] = data[objectColumns].fillna("Unknown") #以分類“Unknown”填充缺失值

import missingno as msno
import matplotlib as mpl
mpl.rcParams['font.sans-serif']=[u'simHei']
mpl.rcParams['axes.unicode_minus']=False
%matplotlib inline
msno.bar(data[objectColumns]) #可視化

<matplotlib.axes._subplots.AxesSubplot at 0x2cacc08aa20>

mapping_dict = {
    "emp_length": {
        "10+ years": 10,
        "9 years": 9,
        "8 years": 8,
        "7 years": 7,
        "6 years": 6,
        "5 years": 5,
        "4 years": 4,
        "3 years": 3,
        "2 years": 2,
        "1 year": 1,
        "< 1 year": 0,
        "n/a": 0
    },
    "grade":{
        "A": 1,
        "B": 2,
        "C": 3,
        "D": 4,
        "E": 5,
        "F": 6,
        "G": 7
    }
}
data = data.replace(mapping_dict) #變量映射

數值類型缺失值

data.select_dtypes(include=[np.number]).isnull().sum().sort_values(ascending=False)

il_util                  883
mths_since_recent_inq    655
mo_sin_old_il_acct       203
mths_since_rcnt_il       203
bc_util                  109
                        ... 
total_cu_tl                0
inq_fi                     0
total_rev_hi_lim           0
total_bc_limit             0
id                         0
Length: 80, dtype: int64

numColumns = data.select_dtypes(include=[np.number]).columns
msno.matrix(data[numColumns]) #缺失值可視化

<matplotlib.axes._subplots.AxesSubplot at 0x2caecfe1160>

data.select_dtypes(include=[np.number])

	id	loan_amnt	funded_amnt	funded_amnt_inv	int_rate	installment	grade	loan_status	dti	fico_range_low	...	num_tl_90g_dpd_24m	num_tl_op_past_12m	pct_tl_nvr_dlq	percent_bc_gt_75	pub_rec_bankruptcies	tax_liens	tot_hi_cred_lim	total_bal_ex_mort	total_bc_limit	total_il_high_credit_limit
18821	163425898	4500	4500	4500	0.1612	158.48	3	1	16.13	705	...	0	2	100.0	28.6	0	0	44700	10872	32800	0.0
61234	161908366	20000	20000	20000	0.2305	564.39	4	0	34.14	735	...	0	0	100.0	33.3	0	0	54349	19572	10400	22349.0
119781	159901427	10000	10000	10000	0.1862	257.32	4	0	27.84	680	...	0	3	100.0	0.0	0	0	69077	48184	9600	49477.0
49201	162292591	21000	21000	21000	0.1430	491.91	3	1	21.82	740	...	0	0	100.0	0.0	0	0	109894	66662	33800	67194.0
53727	162154208	40000	40000	40000	0.0819	814.70	1	0	27.52	700	...	0	0	100.0	50.0	0	0	207370	160985	98000	61725.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
86547	160719957	30000	30000	30000	0.0819	611.03	1	0	5.68	740	...	0	2	100.0	40.0	0	0	361548	46148	94500	0.0
69734	161401437	16000	16000	16000	0.1430	549.18	3	1	13.73	660	...	0	0	90.9	66.7	0	0	21300	15022	7800	6000.0
30947	162968064	1600	1600	1600	0.1102	52.40	2	0	17.32	715	...	0	1	100.0	50.0	0	0	63659	41808	27200	30259.0
29039	163064608	10000	10000	10000	0.1240	334.06	2	0	22.91	680	...	0	2	66.7	0.0	0	0	230024	36479	2900	60846.0
92872	160838177	23000	23000	23000	0.1774	580.81	3	1	0.00	800	...	0	0	100.0	0.0	0	0	85255	0	600	0.0

5973 rows × 80 columns

data.isnull().sum().sum()
mean_cols=data.mean()
data= data.fillna(mean_cols)

目標變量

y=data['loan_status']
x=data.drop(['loan_status'],axis=1)
#使用pandas庫將類別變量編碼
x =pd.get_dummies(x)

n_sample = y.shape[0]
n_pos_sample = y[y == 0].shape[0]
n_neg_sample = y[y == 1].shape[0]
print('樣本個數：{}; 正樣本佔{:.2%}; 負樣本佔{:.2%}'.format(n_sample,
                                                   n_pos_sample / n_sample,
                                                   n_neg_sample / n_sample))
print('特徵維數：', x.shape[1])

樣本個數：5973; 正樣本佔66.97%; 負樣本佔33.03%
特徵維數： 7167

特徵工程

#數據進行分割（訓練數據和測試數據）
from sklearn.model_selection  import train_test_split#測試集和訓練集
x_train1, x_test1, y_train1, y_test1 = train_test_split(x, y, train_size=0.8, random_state=14)

x_train, x_test, y_train, y_test = x_train1, x_test1, y_train1, y_test1
print ("訓練數據集樣本數目：%d, 測試數據集樣本數目：%d" % (x_train.shape[0], x_test.shape[0]))
y_train = y_train.astype(np.int)
y_test = y_test.astype(np.int)

訓練數據集樣本數目：4778, 測試數據集樣本數目：1195

#參數優化
from sklearn.pipeline import Pipeline #管道
from sklearn.model_selection import GridSearchCV #網格搜索交叉驗證，用於選擇最優的參數
from sklearn.preprocessing import MinMaxScaler
from sklearn.decomposition import PCA
pipes =Pipeline([
            ('mms', MinMaxScaler()), ## 歸一化操作
            ('pca', PCA()), ## 降緯
            ('RandomForestClassifier', RandomForestClassifier(criterion='gini'))
        ])
# 參數
#
# estimators = [1,50,100,500]
# depth = [1,2,3,7,15]
parameters = [
    {
    "pca__n_components": [1,2,3,4],
    "RandomForestClassifier__n_estimators":[1,50,100,500],
    "RandomForestClassifier__max_depth":[1,2,3,7,15]
    }
]
#獲取數據
x_train2, x_test2, y_train2, y_test2 = x_train1, x_test1, y_train1, y_test1
gscv = GridSearchCV(pipes, param_grid=parameters)
gscv.fit(x_train2, y_train2)
print ("score值:",gscv.best_score_,"最優參數列表:", gscv.best_params_)

score值: 0.6720405704396591 最優參數列表: {'RandomForestClassifier__max_depth': 7, 'RandomForestClassifier__n_estimators': 500, 'pca__n_components': 4}

#標準化
ss = MinMaxScaler()#分類模型，經常使用的是minmaxscaler歸一化，迴歸模型經常用standardscaler
x_train = ss.fit_transform(x_train, y_train)
x_test = ss.transform(x_test)
x_train.shape

(4778, 7167)

#降維
from sklearn.decomposition import PCA
pca = PCA(n_components=4)
x_train = pca.fit_transform(x_train)
x_test = pca.transform(x_test)
x_train.shape
print(pca.explained_variance_ratio_)

[0.08187674 0.05705152 0.05380546 0.04683824]

#隨機森林模型
from sklearn.ensemble import RandomForestClassifier
forest = RandomForestClassifier(n_estimators=2000, criterion='gini', max_depth=7, random_state=0)
forest.fit(x_train, y_train)#max_depth一般不宜設置過大，把每個模型作爲一個弱分類器

RandomForestClassifier(bootstrap=True, ccp_alpha=0.0, class_weight=None,
                       criterion='gini', max_depth=7, max_features='auto',
                       max_leaf_nodes=None, max_samples=None,
                       min_impurity_decrease=0.0, min_impurity_split=None,
                       min_samples_leaf=1, min_samples_split=2,
                       min_weight_fraction_leaf=0.0, n_estimators=2000,
                       n_jobs=None, oob_score=False, random_state=0, verbose=0,
                       warm_start=False)

#模型效果評估
from sklearn import metrics
from sklearn.metrics import roc_curve, auc
score = forest.score(x_test, y_test)
print ("準確率:%.2f%%" % (score * 100))
#模型預測
y_score = forest.predict(x_test)# prodict_proba輸出概率

準確率:66.78%

# Compute ROC curve and ROC area for each class
import matplotlib.pyplot as plt
fpr,tpr,threshold = roc_curve(y_test, y_score) ###計算真正率和假正率
roc_auc = auc(fpr,tpr) ###計算auc的值
print('auc:%.2f'%(roc_auc))

auc:0.51

plt.figure()
lw = 2
plt.figure(figsize=(10,10))
plt.plot(fpr, tpr, color='darkorange',
         lw=lw, label='ROC curve (area = %0.2f)' % roc_auc) ###假正率爲橫座標，真正率爲縱座標做曲線
plt.plot([0, 1], [0, 1], color='navy', lw=lw, linestyle='--')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver operating characteristic example')
plt.legend(loc="lower right")
plt.show()

<Figure size 432x288 with 0 Axes>

決策樹

#參數優化
from sklearn.tree import DecisionTreeClassifier
pipe = Pipeline([
            ('mms', MinMaxScaler()),
            ('pca', PCA()),
            ('decision', DecisionTreeClassifier(random_state=0))
        ])

# 參數
parameters = {
    "pca__n_components": [0.5,0.99],#設置爲浮點數代表主成分方差所佔最小比例的閾值
    "decision__criterion": ["gini", "entropy"],
    "decision__max_depth": [1,2,3,4,5,6,7,8,9,10]
}
#數據
x_train2, x_test2, y_train2, y_test2 = x_train1, x_test1, y_train1, y_test1
#模型構建：通過網格交叉驗證，尋找最優參數列表， param_grid可選參數列表，cv：進行幾折交叉驗證
gscv = GridSearchCV(pipe, param_grid=parameters,cv=3)
#模型訓練
gscv.fit(x_train2, y_train2)
#算法的最優解
print("最優參數列表:", gscv.best_params_)
print("score值：",gscv.best_score_)

最優參數列表: {'decision__criterion': 'gini', 'decision__max_depth': 4, 'pca__n_components': 0.99}
score值： 0.6917121178186392

#降維
from sklearn.decomposition import PCA
pca = PCA(n_components= 0.99)
x_train = pca.fit_transform(x_train)
x_test = pca.transform(x_test)
x_train.shape
print(pca.explained_variance_ratio_)

[0.34176263 0.23813938 0.22458996 0.19550803]

tree = DecisionTreeClassifier(criterion='gini', max_depth=4)
tree.fit(x_train, y_train) # fit模型訓練
# 模型相關的指標輸出
# print("訓練集上的準確率:%.3f" % tree.score(x_train, y_train))
y_hat = tree.predict(x_test) # 獲取預測值
print("準確率:%.3f" % (np.mean(y_hat == y_test)))

準確率:0.671

# Compute ROC curve and ROC area for each class
import matplotlib.pyplot as plt
fpr,tpr,threshold = roc_curve(y_test, y_hat) ###計算真正率和假正率
roc_auc = auc(fpr,tpr) ###計算auc的值
print('auc:%.2f'%(roc_auc))

auc:0.51

plt.figure()
lw = 2
plt.figure(figsize=(10,10))
plt.plot(fpr, tpr, color='darkorange',
         lw=lw, label='ROC curve (area = %0.2f)' % roc_auc) ###假正率爲橫座標，真正率爲縱座標做曲線
plt.plot([0, 1], [0, 1], color='navy', lw=lw, linestyle='--')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver operating characteristic example')
plt.legend(loc="lower right")
plt.show()

<Figure size 432x288 with 0 Axes>