python信用评分卡（附代码，博主录制）

https://study.163.com/course/introduction.htm?courseId=1005214003&utm_campaign=commission&utm_source=cp-400000000398149&utm_medium=share

## 1. Data Lending Club 2016年Q3数据：https://www.lendingclub.com/info/download-data.action

参考：http://kldavenport.com/lending-club-data-analysis-revisted-with-python/

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt import seaborn as sns %matplotlib inline

df = pd.read_csv("./LoanStats_2016Q3.csv",skiprows=1,low_memory=False)

df.info()

df.head(3)

## 2. Keep what we need

# .ix[row slice, column slice]
df.ix[:4,:7]

df.drop('id',1,inplace=True)
df.drop('member_id',1,inplace=True)

df.int_rate = pd.Series(df.int_rate).str.replace('%', '').astype(float)

df.ix[:4,:7]

### Loan Amount Requested Verus the Funded Amount

print (df.loan_amnt != df.funded_amnt).value_counts()

False 99120 True 4 dtype: int64

df.query('loan_amnt != funded_amnt').head(5)

df.dropna(axis=0, how='all',inplace=True)

df.info()

df.dropna(axis=1, how='all',inplace=True)

df.info()

df.ix[:5,8:15]

### emp_title: employment title

print df.emp_title.value_counts().head()
print df.emp_title.value_counts().tail()
df.emp_title.unique().shape

Teacher 1931 Manager 1701 Owner 990 Supervisor 785 Driver 756 Name: emp_title, dtype: int64 Agent Services Representative 1 Operator Bridge Tunnel 1 Reg Medical Assistant/Referral Spec. 1 Home Health Care 1 rounds cook 1 Name: emp_title, dtype: int64 (37421,)

df.drop(['emp_title'],1, inplace=True)

df.ix[:5,8:15]

### emp_length: employment length

df.emp_length.value_counts()

10+ years 34219 2 years 9066 3 years 7925

df.replace('n/a', np.nan,inplace=True)
df.emp_length.fillna(value=0,inplace=True)
df['emp_length'].replace(to_replace='[^0-9]+', value='', inplace=True, regex=True) df['emp_length'] = df['emp_length'].astype(int)

df.emp_length.value_counts()

10 34219 1 14095 2 9066 3 7925 5 6170 4 6022 0 5922 6 4406 8 4168 9 3922 7 3205 Name: emp_length, dtype: int64 ### verification status:”Indicates if income was verified by LC, not verified, or if the income source was verified”

df.verification_status.value_counts()

Source Verified 40781 Verified 31356 Not Verified 26983 Name: verification_status, dtype: int64 ### Target: Loan Statuses

df.info()

df.columns

Index([u’loan_amnt’, u’funded_amnt’, u’funded_amnt_inv’, u’term’, u’int_rate’, u’installment’, u’grade’, u’sub_grade’, u’emp_length’, u’home_ownership’, … u’num_tl_90g_dpd_24m’, u’num_tl_op_past_12m’, u’pct_tl_nvr_dlq’, u’percent_bc_gt_75’, u’pub_rec_bankruptcies’, u’tax_liens’, u’tot_hi_cred_lim’, u’total_bal_ex_mort’, u’total_bc_limit’, u’total_il_high_credit_limit’], dtype=’object’, length=107)

pd.unique(df['loan_status'].values.ravel())

array([‘Current’, ‘Fully Paid’, ‘Late (31-120 days)’, ‘Charged Off’, ‘Late (16-30 days)’, ‘In Grace Period’, ‘Default’], dtype=object)

for col in df.select_dtypes(include=['object']).columns:print ("Column {} has {} unique instances".format( col, len(df[col].unique())) )

Column term has 2 unique instances Column grade has 7 unique instances Column sub_grade has 35 unique instances Column home_ownership has 4 unique instances Column verification_status has 3 unique instances Column issue_d has 3 unique instances Column loan_status has 7 unique instances Column pymnt_plan has 2 unique instances Column desc has 6 unique instances Column purpose has 13 unique instances Column title has 13 unique instances Column zip_code has 873 unique instances Column addr_state has 50 unique instances Column earliest_cr_line has 614 unique instances Column revol_util has 1087 unique instances Column initial_list_status has 2 unique instances Column last_pymnt_d has 13 unique instances Column next_pymnt_d has 4 unique instances Column last_credit_pull_d has 14 unique instances Column application_type has 3 unique instances Column verification_status_joint has 2 unique instances

# 处理对象类型的缺失，unique
df.select_dtypes(include=['O']).describe().T.\
assign(missing_pct=df.apply(lambda x : (len(x)-x.count())/float(len(x))))

df.revol_util = pd.Series(df.revol_util).str.replace('%', '').astype(float)

# missing_pct
df.drop('desc',1,inplace=True)
df.drop('verification_status_joint',1,inplace=True)

df.drop('zip_code',1,inplace=True)
df.drop('addr_state',1,inplace=True) df.drop('earliest_cr_line',1,inplace=True) df.drop('revol_util',1,inplace=True) df.drop('purpose',1,inplace=True) df.drop('title',1,inplace=True) df.drop('term',1,inplace=True) df.drop('issue_d',1,inplace=True) # df.drop('',1,inplace=True) # 贷后相关的字段 df.drop(['out_prncp','out_prncp_inv','total_pymnt', 'total_pymnt_inv','total_rec_prncp', 'grade', 'sub_grade'] ,1, inplace=True) df.drop(['total_rec_int','total_rec_late_fee', 'recoveries','collection_recovery_fee', 'collection_recovery_fee' ],1, inplace=True) df.drop(['last_pymnt_d','last_pymnt_amnt', 'next_pymnt_d','last_credit_pull_d'],1, inplace=True) df.drop(['policy_code'],1, inplace=True)

df.info()

df.ix[:5,:10]

df.ix[:5,10:21]

print df.columns
print df.head(1).values
df.info()

Index([u’loan_amnt’, u’funded_amnt’, u’funded_amnt_inv’, u’int_rate’, u’installment’, u’emp_length’, u’home_ownership’, u’annual_inc’, u’verification_status’, u’loan_status’, u’pymnt_plan’, u’dti’, u’delinq_2yrs’, u’inq_last_6mths’, u’mths_since_last_delinq’, u’mths_since_last_record’, u’open_acc’, u’pub_rec’, u’revol_bal’, u’total_acc’, u’initial_list_status’, u’collections_12_mths_ex_med’, u’mths_since_last_major_derog’, u’application_type’, u’annual_inc_joint’, u’dti_joint’, u’acc_now_delinq’, u’tot_coll_amt’, u’tot_cur_bal’, u’open_acc_6m’, u’open_il_6m’, u’open_il_12m’, u’open_il_24m’, u’mths_since_rcnt_il’, u’total_bal_il’, u’il_util’, u’open_rv_12m’, u’open_rv_24m’, u’max_bal_bc’, u’all_util’, u’total_rev_hi_lim’, u’inq_fi’, u’total_cu_tl’, u’inq_last_12m’, u’acc_open_past_24mths’, u’avg_cur_bal’, u’bc_open_to_buy’, u’bc_util’, u’chargeoff_within_12_mths’, u’delinq_amnt’, u’mo_sin_old_il_acct’, u’mo_sin_old_rev_tl_op’, u’mo_sin_rcnt_rev_tl_op’, u’mo_sin_rcnt_tl’, u’mort_acc’, u’mths_since_recent_bc’, u’mths_since_recent_bc_dlq’, u’mths_since_recent_inq’, u’mths_since_recent_revol_delinq’, u’num_accts_ever_120_pd’, u’num_actv_bc_tl’, u’num_actv_rev_tl’, u’num_bc_sats’, u’num_bc_tl’, u’num_il_tl’, u’num_op_rev_tl’, u’num_rev_accts’, u’num_rev_tl_bal_gt_0’, u’num_sats’, u’num_tl_120dpd_2m’, u’num_tl_30dpd’, u’num_tl_90g_dpd_24m’, u’num_tl_op_past_12m’, u’pct_tl_nvr_dlq’, u’percent_bc_gt_75’, u’pub_rec_bankruptcies’, u’tax_liens’, u’tot_hi_cred_lim’, u’total_bal_ex_mort’, u’total_bc_limit’, u’total_il_high_credit_limit’], dtype=’object’) [[15000.0 15000.0 15000.0 13.99 512.6 2 ‘RENT’ 55000.0 ‘Not Verified’ ‘Current’ ‘n’ 23.78 1.0 0.0 7.0 nan 22.0 0.0 21345.0 43.0 ‘f’ 0.0 nan ‘INDIVIDUAL’ nan nan 0.0 0.0 140492.0 3.0 10.0 2.0 3.0 11.0 119147.0 101.0 3.0 4.0 14612.0 83.0 39000.0 1.0 6.0 0.0 7.0 6386.0 9645.0 73.1 0.0 0.0 157.0 248.0 4.0 4.0 0.0 4.0 7.0 22.0 7.0 0.0 5.0 9.0 6.0 7.0 25.0 11.0 18.0 9.0 22.0 0.0 0.0 0.0 5.0 100.0 33.3 0.0 0.0 147587.0 140492.0 30200.0 108587.0]]

df.select_dtypes(include=['float']).describe().T.\
assign(missing_pct=df.apply(lambda x : (len(x)-x.count())/float(len(x))))

/Users/ting/anaconda/lib/python2.7/site-packages/numpy/lib/function_base.py:3834: RuntimeWarning: Invalid value encountered in percentile RuntimeWarning)

df.drop('annual_inc_joint',1,inplace=True)
df.drop('dti_joint',1,inplace=True)

df.select_dtypes(include=['int']).describe().T.\
assign(missing_pct=df.apply(lambda x : (len(x)-x.count())/float(len(x))))

Target: Loan Statuses

df['loan_status'].value_counts()
# .plot(kind='bar')

             79445
Fully Paid            13066
Charged Off            2502
Late (31-120 days)     2245
In Grace Period        1407
Late (16-30 days)       454
Default                   1
Name: loan_status, dtype: int64

df.loan_status.replace('Fully Paid', int(1),inplace=True)
df.loan_status.replace('Current', int(1),inplace=True) df.loan_status.replace('Late (16-30 days)', int(0),inplace=True) df.loan_status.replace('Late (31-120 days)', int(0),inplace=True) df.loan_status.replace('Charged Off', np.nan,inplace=True) df.loan_status.replace('In Grace Period', np.nan,inplace=True) df.loan_status.replace('Default', np.nan,inplace=True) # df.loan_status.astype('int') df.loan_status.value_counts()

1.0    92511
0.0     2699
Name: loan_status, dtype: int64

# df.loan_status
df.dropna(subset=['loan_status'],inplace=True)

Highly Correlated Data

cor = df.corr()
cor.loc[:,:] = np.tril(cor, k=-1) # below main lower triangle of an array
cor = cor.stack()
cor[(cor > 0.55) | (cor < -0.55)]

funded_amnt                     loan_amnt                      1.000000
funded_amnt_inv                 loan_amnt                      0.999994funded_amnt                    0.999994
installment                     loan_amnt                      0.953380funded_amnt                    0.953380funded_amnt_inv                0.953293
mths_since_last_delinq          delinq_2yrs                   -0.551275
total_acc                       open_acc                       0.722950
mths_since_last_major_derog     mths_since_last_delinq         0.685642
open_il_24m                     open_il_12m                    0.760219
total_bal_il                    open_il_6m                     0.566551
open_rv_12m                     open_acc_6m                    0.623975
open_rv_24m                     open_rv_12m                    0.774954
max_bal_bc                      revol_bal                      0.551409
all_util                        il_util                        0.594925
total_rev_hi_lim                revol_bal                      0.815351
inq_last_12m                    inq_fi                         0.563011
acc_open_past_24mths            open_acc_6m                    0.553181open_il_24m                    0.570853open_rv_12m                    0.657606open_rv_24m                    0.848964
avg_cur_bal                     tot_cur_bal                    0.828457
bc_open_to_buy                  total_rev_hi_lim               0.626380
bc_util                         all_util                       0.569469
mo_sin_rcnt_tl                  mo_sin_rcnt_rev_tl_op          0.606065
mort_acc                        tot_cur_bal                    0.551198
mths_since_recent_bc            mo_sin_rcnt_rev_tl_op          0.614262
mths_since_recent_bc_dlq        mths_since_last_delinq         0.751613mths_since_last_major_derog    0.553022
mths_since_recent_revol_delinq  mths_since_last_delinq         0.853573...
num_sats                        total_acc                      0.720022num_actv_bc_tl                 0.552957num_actv_rev_tl                0.665429num_bc_sats                    0.630778num_op_rev_tl                  0.826946num_rev_accts                  0.663595num_rev_tl_bal_gt_0            0.668573
num_tl_30dpd                    acc_now_delinq                 0.801444
num_tl_90g_dpd_24m              delinq_2yrs                    0.669267
num_tl_op_past_12m              open_acc_6m                    0.722131open_il_12m                    0.557902open_rv_12m                    0.844841open_rv_24m                    0.660265acc_open_past_24mths           0.774867
pct_tl_nvr_dlq                  num_accts_ever_120_pd         -0.592502
percent_bc_gt_75                bc_util                        0.844108
pub_rec_bankruptcies            pub_rec                        0.580798
tax_liens                       pub_rec                        0.752084
tot_hi_cred_lim                 tot_cur_bal                    0.982693avg_cur_bal                    0.795652mort_acc                       0.560840
total_bal_ex_mort               total_bal_il                   0.902486
total_bc_limit                  max_bal_bc                     0.581536total_rev_hi_lim               0.775151bc_open_to_buy                 0.834159num_bc_sats                    0.633461
total_il_high_credit_limit      open_il_6m                     0.552023total_bal_il                   0.960349num_il_tl                      0.583329total_bal_ex_mort              0.889238
dtype: float64

df.drop(['funded_amnt','funded_amnt_inv', 'installment'], axis=1, inplace=True)

2. Our Model

from sklearn.model_selection import train_test_split
from sklearn.model_selection import GridSearchCV
from sklearn import ensemble from sklearn.preprocessing import OneHotEncoder #https://ljalphabeta.gitbooks.io/python-/content/categorical_data.html

Y = df.loan_status
X = df.drop('loan_status',1,inplace=False)

print Y.shape
print sum(Y)

(95210,)
92511.0

X = pd.get_dummies(X)

print X.columns
print X.head(1).values
X.info()

Index([u'loan_amnt', u'int_rate', u'emp_length', u'annual_inc', u'dti',u'delinq_2yrs', u'inq_last_6mths', u'mths_since_last_delinq',u'mths_since_last_record', u'open_acc', u'pub_rec', u'revol_bal',u'total_acc', u'collections_12_mths_ex_med',u'mths_since_last_major_derog', u'acc_now_delinq', u'tot_coll_amt',u'tot_cur_bal', u'open_acc_6m', u'open_il_6m', u'open_il_12m',u'open_il_24m', u'mths_since_rcnt_il', u'total_bal_il', u'il_util',u'open_rv_12m', u'open_rv_24m', u'max_bal_bc', u'all_util',u'total_rev_hi_lim', u'inq_fi', u'total_cu_tl', u'inq_last_12m',u'acc_open_past_24mths', u'avg_cur_bal', u'bc_open_to_buy', u'bc_util',u'chargeoff_within_12_mths', u'delinq_amnt', u'mo_sin_old_il_acct',u'mo_sin_old_rev_tl_op', u'mo_sin_rcnt_rev_tl_op', u'mo_sin_rcnt_tl',u'mort_acc', u'mths_since_recent_bc', u'mths_since_recent_bc_dlq',u'mths_since_recent_inq', u'mths_since_recent_revol_delinq',u'num_accts_ever_120_pd', u'num_actv_bc_tl', u'num_actv_rev_tl',u'num_bc_sats', u'num_bc_tl', u'num_il_tl', u'num_op_rev_tl',u'num_rev_accts', u'num_rev_tl_bal_gt_0', u'num_sats',u'num_tl_120dpd_2m', u'num_tl_30dpd', u'num_tl_90g_dpd_24m',u'num_tl_op_past_12m', u'pct_tl_nvr_dlq', u'percent_bc_gt_75',u'pub_rec_bankruptcies', u'tax_liens', u'tot_hi_cred_lim',u'total_bal_ex_mort', u'total_bc_limit', u'total_il_high_credit_limit',u'home_ownership_ANY', u'home_ownership_MORTGAGE',u'home_ownership_OWN', u'home_ownership_RENT',u'verification_status_Not Verified',u'verification_status_Source Verified', u'verification_status_Verified',u'pymnt_plan_n', u'pymnt_plan_y', u'initial_list_status_f',u'initial_list_status_w', u'application_type_DIRECT_PAY',u'application_type_INDIVIDUAL', u'application_type_JOINT'],dtype='object')
[[  1.50000000e+04   1.39900000e+01   2.00000000e+00   5.50000000e+042.37800000e+01   1.00000000e+00   0.00000000e+00   7.00000000e+00nan   2.20000000e+01   0.00000000e+00   2.13450000e+044.30000000e+01   0.00000000e+00              nan   0.00000000e+000.00000000e+00   1.40492000e+05   3.00000000e+00   1.00000000e+012.00000000e+00   3.00000000e+00   1.10000000e+01   1.19147000e+051.01000000e+02   3.00000000e+00   4.00000000e+00   1.46120000e+048.30000000e+01   3.90000000e+04   1.00000000e+00   6.00000000e+000.00000000e+00   7.00000000e+00   6.38600000e+03   9.64500000e+037.31000000e+01   0.00000000e+00   0.00000000e+00   1.57000000e+022.48000000e+02   4.00000000e+00   4.00000000e+00   0.00000000e+004.00000000e+00   7.00000000e+00   2.20000000e+01   7.00000000e+000.00000000e+00   5.00000000e+00   9.00000000e+00   6.00000000e+007.00000000e+00   2.50000000e+01   1.10000000e+01   1.80000000e+019.00000000e+00   2.20000000e+01   0.00000000e+00   0.00000000e+000.00000000e+00   5.00000000e+00   1.00000000e+02   3.33000000e+010.00000000e+00   0.00000000e+00   1.47587000e+05   1.40492000e+053.02000000e+04   1.08587000e+05   0.00000000e+00   0.00000000e+000.00000000e+00   1.00000000e+00   1.00000000e+00   0.00000000e+000.00000000e+00   1.00000000e+00   0.00000000e+00   1.00000000e+000.00000000e+00   0.00000000e+00   1.00000000e+00   0.00000000e+00]]
<class 'pandas.core.frame.DataFrame'>
Int64Index: 95210 entries, 0 to 99119
Data columns (total 84 columns):
loan_amnt                              95210 non-null float64
int_rate                               95210 non-null float64
emp_length                             95210 non-null int64
annual_inc                             95210 non-null float64
dti                                    95210 non-null float64
delinq_2yrs                            95210 non-null float64
inq_last_6mths                         95210 non-null float64
mths_since_last_delinq                 51229 non-null float64
mths_since_last_record                 18903 non-null float64
open_acc                               95210 non-null float64
pub_rec                                95210 non-null float64
revol_bal                              95210 non-null float64
total_acc                              95210 non-null float64
collections_12_mths_ex_med             95210 non-null float64
mths_since_last_major_derog            28125 non-null float64
acc_now_delinq                         95210 non-null float64
tot_coll_amt                           95210 non-null float64
tot_cur_bal                            95210 non-null float64
open_acc_6m                            95210 non-null float64
open_il_6m                             95210 non-null float64
open_il_12m                            95210 non-null float64
open_il_24m                            95210 non-null float64
mths_since_rcnt_il                     92660 non-null float64
total_bal_il                           95210 non-null float64
il_util                                82017 non-null float64
open_rv_12m                            95210 non-null float64
open_rv_24m                            95210 non-null float64
max_bal_bc                             95210 non-null float64
all_util                               95204 non-null float64
total_rev_hi_lim                       95210 non-null float64
inq_fi                                 95210 non-null float64
total_cu_tl                            95210 non-null float64
inq_last_12m                           95210 non-null float64
acc_open_past_24mths                   95210 non-null float64
avg_cur_bal                            95210 non-null float64
bc_open_to_buy                         94160 non-null float64
bc_util                                94126 non-null float64
chargeoff_within_12_mths               95210 non-null float64
delinq_amnt                            95210 non-null float64
mo_sin_old_il_acct                     92660 non-null float64
mo_sin_old_rev_tl_op                   95210 non-null float64
mo_sin_rcnt_rev_tl_op                  95210 non-null float64
mo_sin_rcnt_tl                         95210 non-null float64
mort_acc                               95210 non-null float64
mths_since_recent_bc                   94212 non-null float64
mths_since_recent_bc_dlq               24968 non-null float64
mths_since_recent_inq                  85581 non-null float64
mths_since_recent_revol_delinq         35158 non-null float64
num_accts_ever_120_pd                  95210 non-null float64
num_actv_bc_tl                         95210 non-null float64
num_actv_rev_tl                        95210 non-null float64
num_bc_sats                            95210 non-null float64
num_bc_tl                              95210 non-null float64
num_il_tl                              95210 non-null float64
num_op_rev_tl                          95210 non-null float64
num_rev_accts                          95210 non-null float64
num_rev_tl_bal_gt_0                    95210 non-null float64
num_sats                               95210 non-null float64
num_tl_120dpd_2m                       91951 non-null float64
num_tl_30dpd                           95210 non-null float64
num_tl_90g_dpd_24m                     95210 non-null float64
num_tl_op_past_12m                     95210 non-null float64
pct_tl_nvr_dlq                         95210 non-null float64
percent_bc_gt_75                       94156 non-null float64
pub_rec_bankruptcies                   95210 non-null float64
tax_liens                              95210 non-null float64
tot_hi_cred_lim                        95210 non-null float64
total_bal_ex_mort                      95210 non-null float64
total_bc_limit                         95210 non-null float64
total_il_high_credit_limit             95210 non-null float64
home_ownership_ANY                     95210 non-null float64
home_ownership_MORTGAGE                95210 non-null float64
home_ownership_OWN                     95210 non-null float64
home_ownership_RENT                    95210 non-null float64
verification_status_Not Verified       95210 non-null float64
verification_status_Source Verified    95210 non-null float64
verification_status_Verified           95210 non-null float64
pymnt_plan_n                           95210 non-null float64
pymnt_plan_y                           95210 non-null float64
initial_list_status_f                  95210 non-null float64
initial_list_status_w                  95210 non-null float64
application_type_DIRECT_PAY            95210 non-null float64
application_type_INDIVIDUAL            95210 non-null float64
application_type_JOINT                 95210 non-null float64
dtypes: float64(83), int64(1)
memory usage: 61.7 MB

X.fillna(0.0,inplace=True)
X.fillna(0,inplace=True)

Train Data & Test Data

x_train, x_test, y_train, y_test = train_test_split(X, Y, test_size=.3, random_state=123)

print(x_train.shape)
print(y_train.shape)
print(x_test.shape)
print(y_test.shape)

(66647, 84)
(66647,)
(28563, 84)
(28563,)

print y_train.value_counts()
print y_test.value_counts()

1.0    64712
0.0     1935
Name: loan_status, dtype: int64
1.0    27799
0.0      764
Name: loan_status, dtype: int64

Gradient Boosting Regression Tree

# param_grid = {'learning_rate': [0.1, 0.05, 0.02, 0.01],
#               'max_depth': [1,2,3,4],
#               'min_samples_split': [50,100,200,400],
#               'n_estimators': [100,200,400,800]
# } param_grid = {'learning_rate': [0.1], 'max_depth': [2], 'min_samples_split': [50,100], 'n_estimators': [100,200] } # param_grid = {'learning_rate': [0.1], # 'max_depth': [4], # 'min_samples_leaf': [3], # 'max_features': [1.0], # } est = GridSearchCV(ensemble.GradientBoostingRegressor(), param_grid, n_jobs=4, refit=True) est.fit(x_train, y_train) best_params = est.best_params_ print best_params

print best_params

• 1

{'min_samples_split': 100, 'n_estimators': 100, 'learning_rate': 0.1, 'max_depth': 3}

• 1
• 2

%%time
est = ensemble.GradientBoostingRegressor(min_samples_split=50,n_estimators=300,learning_rate=0.1,max_depth=1, random_state=0,loss='ls').\ fit(x_train, y_train)

• 1
• 2
• 3
• 4

CPU times: user 24.2 s, sys: 251 ms, total: 24.4 s
Wall time: 25.6 s

• 1
• 2
• 3

est.score(x_test,y_test)

• 1

0.028311715416075908

• 1
• 2

%%time
est = ensemble.GradientBoostingRegressor(min_samples_split=50,n_estimators=100,learning_rate=0.1,max_depth=2, random_state=0,loss='ls').\ fit(x_train, y_train)

• 1
• 2
• 3
• 4

CPU times: user 20 s, sys: 272 ms, total: 20.3 s
Wall time: 21.6 s

• 1
• 2
• 3

est.score(x_test,y_test)

• 1

0.029210266192750467

• 1
• 2

def compute_ks(data):sorted_list = data.sort_values(['predict'], ascending=[True]) total_bad = sorted_list['label'].sum(axis=None, skipna=None, level=None, numeric_only=None) / 3 total_good = sorted_list.shape[0] - total_bad # print "total_bad = ", total_bad # print "total_good = ", total_good max_ks = 0.0 good_count = 0.0 bad_count = 0.0 for index, row in sorted_list.iterrows(): if row['label'] == 3: bad_count += 1.0 else: good_count += 1.0 val = bad_count/total_bad - good_count/total_good max_ks = max(max_ks, val) return max_ks

test_pd = pd.DataFrame()
test_pd['predict'] = est.predict(x_test)
test_pd['label'] = y_test
# df['predict'] = est.predict(x_test)
print compute_ks(test_pd[['label','predict']])

0.0

# Top Ten
feature_importance = est.feature_importances_
feature_importance = 100.0 * (feature_importance / feature_importance.max())indices = np.argsort(feature_importance)[-10:]
plt.barh(np.arange(10), feature_importance[indices],color='dodgerblue',alpha=.4) plt.yticks(np.arange(10 + 0.25), np.array(X.columns)[indices]) _ = plt.xlabel('Relative importance'), plt.title('Top Ten Important Variables')

Other Model

import xgboost as xgb
from sklearn.ensemble import ExtraTreesRegressor, RandomForestRegressor

• 1
• 2

# XGBoost
clf2 = xgb.XGBClassifier(n_estimators=50, max_depth=1, learning_rate=0.01, subsample=0.8, colsample_bytree=0.3,scale_pos_weight=3.0, silent=True, nthread=-1, seed=0, missing=None,objective='binary:logistic', reg_alpha=1, reg_lambda=1, gamma=0, min_child_weight=1, max_delta_step=0,base_score=0.5) clf2.fit(x_train, y_train) print clf2.score(x_test, y_test) test_pd2 = pd.DataFrame() test_pd2['predict'] = clf2.predict(x_test) test_pd2['label'] = y_test print compute_ks(test_pd[['label','predict']]) print clf2.feature_importances_ # Top Ten feature_importance = clf2.feature_importances_ feature_importance = 100.0 * (feature_importance / feature_importance.max()) indices = np.argsort(feature_importance)[-10:] plt.barh(np.arange(10), feature_importance[indices],color='dodgerblue',alpha=.4) plt.yticks(np.arange(10 + 0.25), np.array(X.columns)[indices]) _ = plt.xlabel('Relative importance'), plt.title('Top Ten Important Variables')

0.973252109372
0.0
[ 0.          0.30769232  0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.051282050.          0.          0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.          0.0.          0.          0.          0.          0.          0.          0.0.05128205  0.30769232  0.2820513   0.          0.          0.          0.0.        ]

# RFR
clf3 = RandomForestRegressor(n_jobs=-1, max_depth=10,random_state=0)
clf3.fit(x_train, y_train)
print clf3.score(x_test, y_test) test_pd3 = pd.DataFrame() test_pd3['predict'] = clf3.predict(x_test) test_pd3['label'] = y_test print compute_ks(test_pd[['label','predict']]) print clf3.feature_importances_ # Top Ten feature_importance = clf3.feature_importances_ feature_importance = 100.0 * (feature_importance / feature_importance.max()) indices = np.argsort(feature_importance)[-10:] plt.barh(np.arange(10), feature_importance[indices],color='dodgerblue',alpha=.4) plt.yticks(np.arange(10 + 0.25), np.array(X.columns)[indices]) _ = plt.xlabel('Relative importance'), plt.title('Top Ten Important Variables')

0.0148713087517
0.0
[ 0.02588781  0.10778862  0.00734994  0.02090219  0.02231172  0.007780160.00556834  0.01097013  0.00734689  0.0017027   0.00622544  0.011408430.00530896  0.00031185  0.01135318  0.          0.01488991  0.018405590.00585621  0.00652523  0.0066759   0.00727607  0.00955013  0.010046720.01785864  0.00855197  0.00985739  0.01477432  0.02184904  0.018161840.00878854  0.02078236  0.01310288  0.00844302  0.01596395  0.018251960.01817367  0.00297759  0.00084823  0.02808718  0.02917066  0.008970340.01139324  0.01532409  0.01467681  0.0032855   0.01066291  0.005816610.00955357  0.00417743  0.01333577  0.00489264  0.0128039   0.013401950.01286394  0.01619219  0.00395603  0.00508973  0.          0.002347570.00378329  0.00502684  0.01732834  0.01178674  0.00030035  0.011895090.00942532  0.00841645  0.01571355  0.00288054  0.          0.00116670.00106548  0.00488734  0.          0.00200132  0.00062765  0.041308730.10076558  0.00022293  0.00165858  0.00308408  0.0008255   0.        ]

# XTR
clf4 = ExtraTreesRegressor(n_jobs=-1, max_depth=10,random_state=0)
clf4.fit(x_train, y_train)
print clf4.score(x_test, y_test) test_pd4 = pd.DataFrame() test_pd4['predict'] = clf4.predict(x_test) test_pd4['label'] = y_test print compute_ks(test_pd[['label','predict']]) print clf4.feature_importances_ # Top Ten feature_importance = clf4.feature_importances_ feature_importance = 100.0 * (feature_importance / feature_importance.max()) indices = np.argsort(feature_importance)[-10:] plt.barh(np.arange(10), feature_importance[indices],color='dodgerblue',alpha=.4) plt.yticks(np.arange(10 + 0.25), np.array(X.columns)[indices]) _ = plt.xlabel('Relative importance'), plt.title('Top Ten Important Variables')

0.020808034579
0.0
[ 0.00950112  0.17496689  0.00476969  0.00538677  0.00898343  0.016048850.0139889   0.00605683  0.0042762   0.00358536  0.0144985   0.009151890.00643305  0.00637134  0.0050764   0.00218012  0.00925068  0.003633390.00988441  0.00645297  0.00662444  0.00934969  0.00739012  0.006355920.00633908  0.00923972  0.01263829  0.01190224  0.00914159  0.004021440.00917841  0.01456563  0.01161155  0.01097394  0.00506868  0.007721590.00560163  0.01132941  0.00172528  0.0085601   0.01282485  0.009706290.00956066  0.00731205  0.02087289  0.00430205  0.0062769   0.007656930.00922104  0.00296456  0.00563208  0.00459181  0.0133819   0.005482080.00450864  0.0132415   0.00677772  0.00509891  0.00108962  0.005784480.00934323  0.00715127  0.01078137  0.00855071  0.00695096  0.014889930.00317962  0.00485367  0.00476553  0.00509674  0.          0.007336540.00097223  0.00380448  0.00534715  0.00356893  0.0128526   0.119445380.11758343  0.00195945  0.00225379  0.00243429  0.0007562   0.        ]

作业：

1. feature-engineering

2. stacking

3. 画出ROC曲线和KS曲线对比

# 特征工程方法1：histogram
def get_histogram_features(full_dataset): def extract_histogram(x): count, _ = np.histogram(x, bins=[0, 10, 100, 1000, 10000, 100000, 1000000, 9000000]) return count column_names = ["hist_{}".format(i) for i in range(8)] hist = full_dataset.apply(lambda row: pd.Series(extract_histogram(row)), axis=1) hist.columns= column_names RETURN hist # 特征工程方法2：quantile q = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] column_names = ["quantile_{}".format(i) for i in q] # print pd.DataFrame(train_x) quantile = pd.DataFrame(x_train).quantile(q=q, axis=1).T quantile.columns = column_names # 特征工程方法3：cumsum def get_cumsum_features(all_features): column_names = ["cumsum_{}".format(i) for i in range(len(all_features))] cumsum = full_dataset[all_features].cumsum(axis=1) cumsum.columns = column_names return cumsum # 特征工程方法4：特征归一化 from sklearn.preprocessing import MinMaxScaler Scaler = MinMaxScaler() x_train_normal = Scaler.fit_transform(x_train_normal)