个贷违约
Contents
1.9. 个贷违约#
PDF请访问个贷违约预测.pdf。
1.9.1. 使用autoML模型#
# # 用autoML 来找最佳的metric - path 只是model信息保存的文件路径
# model = TabularPredictor(label="subscribe", eval_metric='f1', path="model_simple")
# # 获取模型各个feature的重要性
# model.feature_metadata_in.get_features()
# model.feature_importance(train)
1.9.2. 特征工程：考虑每个feature实际的重要性#
1.9.2.1. 一种numeric格式转换成categoric的格式的方式#
for a list of pred, makes it be a human-readable labels and input to a dataframe as a column
y_pred = [1 if x >= 0.5 else 0 for x in y_pred]
subscribe_map = {1.0:”yes”, 0.0:”no”}
result[‘subscribe’] = [subscribe_map[x] for x in y_pred]
# 最后一位按照年加进去
def f(x):
try:
x1, x2, x3 = x.split('-')
if x1 == '2022':
x = str(int(x3) + 2000) + '/' + x2 + '/' + '1'
except Exception:
x == x
return x
train_public['earlies_credit_mon'] = train_public['earlies_credit_mon'].apply(f)
test_public['earlies_credit_mon'] = test_public['earlies_credit_mon'].apply(f)
def process_cat(data):
# class employer_type industry work_year
# 先看class有多少类
class_map = {'A':6, 'B':5, 'C':4, 'D':3, 'E':2, 'F':1, 'G':0}
yr_map = {'< 1 year':0, '1 years':1, '2 years':2, '3 years':3, '4 years':4, '5 years':5, '6 years':6, '7 years':7, '8 years':8, '9 years':9, '10+ years':10}
industry_map = {'金融业':0, '房地产业':0, '电力、热力生产供应业':1, '制造业':1, '建筑业':1, '交通运输、仓储和邮政业':1, '农、林、牧、渔业':2, '采矿业':2, '批发和零售业':3, '住宿和餐饮业':3, '文化和体育业':3, '公共服务、社会组织':4, '国际组织':4, '信息传输、软件和信息技术服务业':5}
emp_map = {'普通企业':0, '政府机构':1, '世界五百强':3, '高等教育机构':1, '上市企业':2, '幼教与中小学校':1}
'''
单纯的用于表示cat的numeric变成程度关联
金融业 122758
电力、热力生产供应业 92072
公共服务、社会组织 77326
住宿和餐饮业 68422
文化和体育业 61548
信息传输、软件和信息技术服务业 61023
建筑业 53330
房地产业 45733
交通运输、仓储和邮政业 38300
采矿业 38092
农、林、牧、渔业 38091
国际组织 22909
批发和零售业 22738
制造业 22658
普通企业 347404
政府机构 197443
幼教与中小学校 76547
上市企业 76425
世界五百强 41361
高等教育机构 25820
'''
# 其他几个类用labelencoder
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
# data['employer_type'] = le.fit_transform(data['employer_type'])
data['industry'] = le.fit_transform(data['industry'])
# data['industry'] = data['industry'].map(industry_map)
data['class'] = data['class'].map(class_map)
data['work_year'] = data['work_year'].map(yr_map)
data['employer_type'] = data['employer_type'].map(emp_map)
# 看到有时间，就先处理时间类型
import datetime
data['earlies_credit_mon'] = pd.to_datetime(data['earlies_credit_mon'])
# 时间多尺度, 都是月度时间，用value_counts()可以看到日期日的话都是1
data['earlies_credit_mon_yr'] = data['earlies_credit_mon'].dt.year
data['earlies_credit_mon_mt'] = data['earlies_credit_mon'].dt.month
# 将时间转换成时间区间
# 首先设置basetime
basetime1 = data['earlies_credit_mon'].min()
data['earlies_credit_mon_diffs'] = data['earlies_credit_mon'].apply(lambda x:x-basetime1).dt.days
# train['issue_date_diffs']
data['issue_date'] = pd.to_datetime(data['issue_date'])
# 时间多尺度, 都是月度时间，用value_counts()可以看到日期日的话都是1
data['issue_year'] = data['issue_date'].dt.year
data['issue_month'] = data['issue_date'].dt.month
# 将时间转换成时间区间
# 首先设置basetime
basetime2 = data['issue_date'].min()
data['issue_date_diffs'] = data['issue_date'].apply(lambda x:x-basetime2).dt.days
# train['issue_date_diffs']
# drop多余的列
# data.drop(['issue_date', 'earlies_credit_mon'], axis=1, inplace=True)
return data
# 如果打印出来这两个table， 发现有些feature是一样的但是名字不一样
train_public = train_public.rename(columns={'isDefault':'is_default'})
# 查找数据集ab
common_cols = []
for col in train_public.columns:
if col in train_internet:
common_cols.append(col)
len(common_cols)
36
print(set(train_internet.columns) - set(common_cols))
{'f5', 'house_loan_status', 'marriage', 'sub_class', 'offsprings', 'work_type'}
import numpy as np
train_internet = train_internet.drop(['work_type', 'marriage', 'sub_class', 'f5', 'house_loan_status', 'offsprings'], axis=1)
for col in ['known_outstanding_loan', 'app_type', 'known_dero']:
train_internet[col] = np.nan
# 增加 is_internet 字段
train_internet['is_internet'] = True
train_public['is_internet'] = False
test_public['is_internet'] = False
# 拼接所有数据
df = pd.concat([train_public, test_public, train_internet]).reset_index(drop=True)
df = process_cat(df)
# ---------- #
'''['loan_id', 'user_id', 'total_loan', 'year_of_loan', 'interest',
'monthly_payment', 'class', 'employer_type', 'industry', 'work_year',
'house_exist', 'censor_status', 'issue_date', 'use', 'post_code',
'region', 'debt_loan_ratio', 'del_in_18month', 'scoring_low',
'scoring_high', 'pub_dero_bankrup', 'early_return',
'early_return_amount', 'early_return_amount_3mon', 'recircle_b',
'recircle_u', 'initial_list_status', 'earlies_credit_mon', 'title',
'policy_code', 'f0', 'f1', 'f2', 'f3', 'f4', 'is_default',
'known_outstanding_loan', 'app_type', 'known_dero', 'is_internet'],
'''
df['monthly_loan_rb'] = df['monthly_payment'] * 12 * df['year_of_loan'] / (df['recircle_b'] + 0.1)
# 单位循环额度利用率
df['rurb'] = df['recircle_u'] / (df['recircle_b'] + 0.1)
# debt_loan_ratio 债务收入比
# 对于 <0 的异常值进行处理 ==》 转换为正数
df['debt_loan_ratio'] = np.abs(df['debt_loan_ratio']).fillna(df['debt_loan_ratio'].mean())
# recircle_b_total_loan 信贷额度周转率
df['recircle_b_total_loan'] = df['recircle_b'] / (df['total_loan'] + 0.1)
# recircle_b_monthly_pmt 分期付款金额
df['recircle_bmp'] = df['recircle_b'] / (df['monthly_payment'] + 0.1)
# 空缺情况, =1 统计有多少空列 =0 统计有多少空行
df['nulls'] = df.isnull().sum(axis=1)
# 平均每年贷款金额
df['avg_loan'] = df['total_loan'] / df['year_of_loan']
# 平均每年贷款利率
df['mean_interest'] = df['interest'] / df['year_of_loan']
# 总计还款金额
df['all_monthly_pmt'] = df['monthly_payment'] * df['year_of_loan']
# rest_known_zero 剩余公开记录的数量
df['rest_known_dero'] = df['known_dero'] - df['pub_dero_bankrup']
# 除信贷周转余额外，剩余贷款数额
df['rest_loan_recircle_b'] = df['total_loan'] - df['recircle_b']
1.9.3. 数据分箱#
# 不适合太多， 稳定性， 每个分箱的样本数 > 5%
# df.groupby(['industry'])['is_default'].mean()
bin_number = 12
label_list = list(range(bin_number))
df['total_loan_bin'] = pd.qcut(df['total_loan'], bin_number, labels=label_list, duplicates='drop')
df_internet_f = df[df['is_internet'] == False].copy()
df_internet_t = df[df['is_internet'] == True].copy()
db_train = df_internet_f[df_internet_f['is_default'].notnull()]
db_test = df_internet_f[df_internet_f['is_default'].isnull()]
# df.select_dtypes(include='O')
1.9.4. 模型训练#
data = pd.concat([db_train, df_internet_t])
from autogluon.tabular import TabularPredictor
model_autogluon = TabularPredictor(label='is_default',eval_metric='roc_auc', path="./model")
model_autogluon.fit(data, time_limit=500, hyperparameters = {
# 'GBM': {'num_boost_round': 5000},
'CAT': {'iterations': 2000},
# 'XT': {'n_estimators': 300},
})
model_autogluon.leaderboard()
Warning: path already exists! This predictor may overwrite an existing predictor! path="./model"
Beginning AutoGluon training ... Time limit = 500s
AutoGluon will save models to "./model\"
AutoGluon Version: 0.4.0
Python Version: 3.8.3
Operating System: Windows
Train Data Rows: 760000
Train Data Columns: 56
Label Column: is_default
Preprocessing data ...
AutoGluon infers your prediction problem is: 'binary' (because only two unique label-values observed).
2 unique label values: [0.0, 1.0]
If 'binary' is not the correct problem_type, please manually specify the problem_type parameter during predictor init (You may specify problem_type as one of: ['binary', 'multiclass', 'regression'])
Selected class <--> label mapping: class 1 = 1, class 0 = 0
Using Feature Generators to preprocess the data ...
Fitting AutoMLPipelineFeatureGenerator...
Available Memory: 5099.24 MB
Train Data (Original) Memory Usage: 326.8 MB (6.4% of available memory)
Warning: Data size prior to feature transformation consumes 6.4% of available memory. Consider increasing memory or subsampling the data to avoid instability.
Inferring data type of each feature based on column values. Set feature_metadata_in to manually specify special dtypes of the features.
Stage 1 Generators:
Fitting AsTypeFeatureGenerator...
Note: Converting 3 features to boolean dtype as they only contain 2 unique values.
Stage 2 Generators:
Fitting FillNaFeatureGenerator...
Stage 3 Generators:
Fitting IdentityFeatureGenerator...
Fitting CategoryFeatureGenerator...
Fitting CategoryMemoryMinimizeFeatureGenerator...
Fitting DatetimeFeatureGenerator...
Stage 4 Generators:
Fitting DropUniqueFeatureGenerator...
Useless Original Features (Count: 1): ['policy_code']
These features carry no predictive signal and should be manually investigated.
This is typically a feature which has the same value for all rows.
These features do not need to be present at inference time.
Types of features in original data (raw dtype, special dtypes):
('bool', []) : 1 | ['is_internet']
('category', []) : 1 | ['total_loan_bin']
('datetime', []) : 2 | ['issue_date', 'earlies_credit_mon']
('float', []) : 31 | ['total_loan', 'interest', 'monthly_payment', 'work_year', 'post_code', ...]
('int', []) : 20 | ['loan_id', 'user_id', 'year_of_loan', 'class', 'employer_type', ...]
Types of features in processed data (raw dtype, special dtypes):
('category', []) : 1 | ['total_loan_bin']
('float', []) : 31 | ['total_loan', 'interest', 'monthly_payment', 'work_year', 'post_code', ...]
('int', []) : 18 | ['loan_id', 'user_id', 'class', 'employer_type', 'industry', ...]
('int', ['bool']) : 3 | ['year_of_loan', 'initial_list_status', 'is_internet']
('int', ['datetime_as_int']) : 8 | ['issue_date', 'issue_date.year', 'issue_date.month', 'issue_date.dayofweek', 'earlies_credit_mon', ...]
5.7s = Fit runtime
55 features in original data used to generate 61 features in processed data.
Train Data (Processed) Memory Usage: 328.32 MB (6.2% of available memory)
Data preprocessing and feature engineering runtime = 6.77s ...
AutoGluon will gauge predictive performance using evaluation metric: 'roc_auc'
This metric expects predicted probabilities rather than predicted class labels, so you'll need to use predict_proba() instead of predict()
To change this, specify the eval_metric parameter of Predictor()
Automatically generating train/validation split with holdout_frac=0.01, Train Rows: 752400, Val Rows: 7600
Fitting 1 L1 models ...
Fitting model: CatBoost ... Training model for up to 493.23s of the 493.23s of remaining time.
Warning: Potentially not enough memory to safely train CatBoost model, roughly requires: 2.483 GB, but only 4.867 GB is available...
Ran out of time, early stopping on iteration 1499.
0.7969 = Validation score (roc_auc)
493.63s = Training runtime
0.02s = Validation runtime
Fitting model: WeightedEnsemble_L2 ... Training model for up to 360.0s of the -0.48s of remaining time.
0.7969 = Validation score (roc_auc)
0.01s = Training runtime
0.0s = Validation runtime
AutoGluon training complete, total runtime = 501.64s ... Best model: "WeightedEnsemble_L2"
TabularPredictor saved. To load, use: predictor = TabularPredictor.load("./model\")
model score_val pred_time_val fit_time pred_time_val_marginal fit_time_marginal stack_level can_infer fit_order
0 CatBoost 0.796899 0.015957 493.625306 0.015957 493.625306 1 True 1
1 WeightedEnsemble_L2 0.796899 0.018947 493.631290 0.002990 0.005985 2 True 2
| model | score_val | pred_time_val | fit_time | pred_time_val_marginal | fit_time_marginal | stack_level | can_infer | fit_order | |
|---|---|---|---|---|---|---|---|---|---|
| 0 | CatBoost | 0.796899 | 0.015957 | 493.625306 | 0.015957 | 493.625306 | 1 | True | 1 |
| 1 | WeightedEnsemble_L2 | 0.796899 | 0.018947 | 493.631290 | 0.002990 | 0.005985 | 2 | True | 2 |
# predict
y_pred = model_autogluon.predict_proba(db_test)[1]
result = pd.DataFrame(columns=['id', 'isDefault'])
result['id'] = db_test['loan_id']
result['isDefault'] = y_pred
result.to_csv('/Users/19723/Desktop/baseline_autogluon6.csv')
data.drop(['is_default'], axis=1)
import lightgbm as lgb
import xgboost as xgb
from catboost import CatBoostClassifier
# model_lgb = lgb.LGBMClassifier(n_estimators=500)
xgb_lgb = xgb.XGBClassifier(learning_rate=0.1, n_estimators=150, gamma=0,subsample=0.8, colsample_bytree=0.8, max_depth=7, random_state=2022, enable_categorical=True)
# model_Cat = CatBoostClassifier(iterations=5000, depth=7, learning_rate=0.001, loss_function='Logloss', eval_metric='AUC', logging_level='Verbose', metric_period=50)
# 模型训练
labels = data['is_default']
xgb_lgb.fit(data.drop(['is_default'], axis=1), labels)
y_pred = xgb_lgb.predict_proba(db_test)
result = pd.DataFrame(columns=['id', 'isDefault'])
result['id'] = test['loan_id']
result['isDefault'] = y_pred[:,1]
result