本文围绕天池贷款违约预测赛题，介绍零基础入门金融风控的实现过程。使用超过120万条贷款记录数据，对比多种数据不平衡处理方法，以CatBoost为基分类器，经5折交叉验证，结合过采样的模型线上表现较好（AUC 0.7347），SMOTE和BalanceCascade效果欠佳，为相关任务提供参考。

零基础入门金融风控-贷款违约预测

该赛题来自 TIANCHI 零基础入门系列，仅用作学习交流

该数据来自某信贷平台的贷款记录，总数据量超过120w，包含47列变量信息，其中15列为匿名变量。为了保证比赛的公平性，将会从中抽取80万条作为训练集，20万条作为测试集A，20万条作为测试集B，同时会对employmentTitle、purpose、postCode和title等信息进行脱敏。

免费影视、动漫、音乐、游戏、小说资源长期稳定更新！ 👉 点此立即查看 👈

字段表

参考：

[1] https://github.com/caozichuan/TianChi_loadDefault/blob/main/GitHub_loadDefault/model/xgb_github.ipynb

[2] https://tianchi.aliyun.com/notebook-ai/detail?spm=5176.21852664.0.0.4f70379c1gFeCt&postId=129321

[3] https://imbalanced-ensemble.readthedocs.io/en/latest/index.html

[4] https://blog.csdn.net/qq_31367861/article/details/111145816?spm=5176.21852664.0.0.22686e12UZfkmZ

文献4方法线上提交分数为0.7391

登录后复制

       

[5] https://blog.csdn.net/qq_44694861/article/details/109753004?spm=5176.21852664.0.0.667f4288swn2OQ

本项目主要是对金融风控中常见的数据不平衡方法进行对比，以及对 self-paced-ensemble 包进行学习

使用的方法如下：

无处理SMOTE算法

SMOTE（Synthetic Minority Oversampling Technique），合成少数类过采样技术．它是基于随机过采样算法的一种改进方案，为了克服随机过采样算法泛化能力差的缺点，SMOTE算法的对少数类样本近邻进行采样，根据少数类样本人工合成新样本添加到数据集中。

SPE算法

Liu Z , Cao W , Gao Z , et al. Self-paced Ensemble for Highly Imbalanced Massive Data Classification[J]. 2019.

过采样算法

重复正样本，使得正样本与负样本比例接近 1 : 1

欠采样算法

从数量多的样本里面随机选择样本进行抛弃,为了避免随机性影响，我们独立训练多个模型进行简单平均

BalanceCascade算法

Liu, X. Y., Wu, J., & Zhou, Z. H. “Exploratory undersampling for class-imbalance learning.” IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 39.2 (2008): 539-550

注意事项

运行主要有两部分，数据处理 和 模型选择 ,根据注释部分，自己进行数据处理和算法组合。默认CBT模型+过采样算法。In [1]

#!pip install pandas --user#!pip install lightgbm --user#!pip install xgboost --user!pip install catboost --user!pip install joblib --user!pip install scikit-learn --user!pip install imblearn --user!pip install imbalanced-ensemble --user!pip self-paced-ensemble --user

登录后复制    In [2]

import pandas as pdimport datetimeimport warningswarnings.filterwarnings('ignore')from sklearn.model_selection import StratifiedKFold#warnings.filterwarnings('ignore')#%matplotlib inlinefrom sklearn.metrics import roc_auc_score## 数据降维处理的from sklearn.model_selection import train_test_split  from catboost import CatBoostClassifierimport imbalanced_ensemble as imbensfrom collections import Counterfrom imbalanced_ensemble.sampler.over_sampling import     SMOTE import numpy as nptrain=pd.read_csv("./data/data53042/train.csv")testA=pd.read_csv("./data/data53042/testA.csv")numerical_fea = list(train.select_dtypes(exclude=['object']).columns)category_fea = list(filter(lambda x: x not in numerical_fea,list(train.columns)))label = 'isDefault'numerical_fea.remove(label)#按照中位数填充数值型特征train[numerical_fea] = train[numerical_fea].fillna(train[numerical_fea].median())testA[numerical_fea] = testA[numerical_fea].fillna(testA[numerical_fea].median())#按照众数填充类别型特征train[category_fea] = train[category_fea].fillna(train[category_fea].mode())testA[category_fea] = testA[category_fea].fillna(testA[category_fea].mode())def make_fea(data):    data['issueDate'] = pd.to_datetime(data['issueDate'],format='%Y-%m-%d')    startdate = datetime.datetime.strptime('2007-06-01', '%Y-%m-%d')    #构造时间特征    data['issueDateDT'] = data['issueDate'].apply(lambda x: x-startdate).dt.days    data['grade'] = data['grade'].map({'A':1,'B':2,'C':3,'D':4,'E':5,'F':6,'G':7})    data['subGrade'] = data['subGrade'].map({'E2':1,'D2':2,'D3':3,'A4':4,'C2':5,'A5':6,'C3':7,'B4':8,'B5':9,'E5':10,        'D4':11,'B3':12,'B2':13,'D1':14,'E1':15,'C5':16,'C1':17,'A2':18,'A3':19,'B1':20,        'E3':21,'F1':22,'C4':23,'A1':24,'D5':25,'F2':26,'E4':27,'F3':28,'G2':29,'F5':30,        'G3':31,'G1':32,'F4':33,'G4':34,'G5':35})    data = pd.get_dummies(data, columns=['subGrade', 'homeOwnership', 'verificationStatus', 'purpose', 'regionCode'], drop_first=True)    data['employmentLength'] = data['employmentLength'].map({'NaN':-1,'1 year':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,'< 1 year':0})    data['earliesCreditLine'] = data['earliesCreditLine'].apply(lambda s: int(s[-4:]))    for item in ['n0','n1','n2','n2.1','n4','n5','n6','n7','n8','n9','n10','n11','n12','n13','n14']:        data['grade_to_mean_' + item] = data['grade'] / data.groupby([item])['grade'].transform('mean')        data['grade_to_std_' + item] = data['grade'] / data.groupby([item])['grade'].transform('std')    data['n15']=data['n8']*data['n10']    return data# 特征工程 简单处理train = make_fea(train)testA = make_fea(testA)print(train.shape)print(testA.shape)print("数据预处理完成!")  sub=testA[['id']].copy()sub['isDefault']=0testA=testA.drop(['id','issueDate'],axis=1)data_x=train.drop(['isDefault','id','issueDate'],axis=1)data_y=train[['isDefault']].copy()col=['grade','subGrade','employmentTitle','homeOwnership','verificationStatus','purpose','postCode','regionCode',     'initialListStatus','applicationType','policyCode']for i in data_x.columns:    if i in col:        data_x[i] = data_x[i].astype('str')for i in testA.columns:    if i in col:        testA[i] = testA[i].astype('str')answers = []mean_score = 0n_folds = 5sk = StratifiedKFold(n_splits=n_folds, shuffle=True, random_state=2019)print(data_x.shape)print(testA.shape)# 在测试集中有部分特征训练集中没有，进行处理COLNAME = data_x.columnstestA = testA[COLNAME]# 处理INF NAN ，替换成 -1 注意：np.inf和-np.inf是不同的两个值，这里经常遇到忘记处理后者引发错误data_x.replace([np.inf, -np.inf, np.nan], -1, inplace=True)testA.replace([np.inf, -np.inf, np.nan], -1, inplace=True)#  -----重复正样本，因为原始数据比例约为3：1 因此对正样本重复三次拼接到原始数据中-----a = [i for i, x in enumerate(data_y.values) if x[0]==1]all_1 = data_x.iloc[a]all_y = pd.DataFrame(np.ones(all_1.shape[0],dtype=int),columns=['isDefault'])print(all_y.shape)print(data_x.shape,data_y.shape)for i in range(3):    data_y= pd.concat([data_y,all_y],axis=0)    data_x = pd.concat([data_x,all_1],axis=0)    print(data_x.shape,data_y.shape)data_x.reset_index(drop=True,inplace =True)data_y.reset_index(drop=True,inplace =True)'''# -----使用SMOTE算法对负样本进行扩充-----sm = SMOTE(random_state=42)data_x, data_y = sm.fit_resample(data_x, data_y)print('Resampled dataset' , data_y.value_counts())'''

登录后复制    

基础模型定义及训练

使用CatBoost分类模型作为基准模型

In [ ]

model=CatBoostClassifier(            loss_function="Logloss",            eval_metric="AUC",            task_type="CPU",            learning_rate=0.1,            iterations=500,            random_seed=2020,            verbose=500,            depth=7)for train, test in sk.split(data_x, data_y):    x_train = data_x.iloc[train]    y_train = data_y.iloc[train]    x_test = data_x.iloc[test]    y_test = data_y.iloc[test]    '''    # -----使用SPE 算法对不平衡数据进行处理-----    clf = imbens.ensemble.SelfPacedEnsembleClassifier(        base_estimator = model,#基准分类模型 可以自定义，需要模型包含fit等方法，具体信息请查看参考文献3        n_estimators = 5,        random_state=49    ).fit(x_train, y_train)    '''    # -----使用BalanceCascade 算法对不平衡数据进行处理-----    '''    clf = imbens.ensemble.BalanceCascadeClassifier(                base_estimator = model,#基准分类模型 可以自定义，需要模型包含fit等方法，具体信息请查看参考文献3        n_estimators = 5,        random_state=49    ).fit(x_train, y_train)    '''    # 基准分类模型    clf = model.fit(x_train,y_train)#, eval_set=(x_test,y_test),cat_features=col)    yy_pred_valid=clf.predict_proba(x_test)[:,-1]    print('cat验证的auc:{}'.format(roc_auc_score(y_test, yy_pred_valid)))    mean_score += roc_auc_score(y_test, yy_pred_valid) / n_folds    y_pred_valid = clf.predict_proba(testA)[:,-1]    answers.append(y_pred_valid)print('mean valAuc:{}'.format(mean_score))cat_pre=sum(answers)/n_foldssub['isDefault']=cat_presub.to_csv('金融预测.csv',index=False)

登录后复制    

欠采样模型训练

In [ ]

model=CatBoostClassifier(            loss_function="Logloss",            eval_metric="AUC",            task_type="CPU",            learning_rate=0.1,            iterations=500,            random_seed=2020,            verbose=500,            depth=7)#----- 欠采样模型 -----a = [i for i, x in enumerate(data_y.values) if x[0]==1]all_1 = data_x.iloc[a]all_y1 = pd.DataFrame(np.ones(all_1.shape[0]),columns=['isDefault'])all_y0 = pd.DataFrame(np.ones(all_y.shape[0]),columns=['isDefault'])print(all_y1.shape)#for i in range(4): # 数据比1：4+i= 0all_0 = data_x.iloc[i*all_y1.shape[0]:(i+1)*all_y1.shape[0],:]data_y_= pd.concat([all_y1,all_y0],axis=0)data_x_ = pd.concat([all_1,all_0],axis=0)print(data_x_.shape,data_y_.shape,all_0.shape)for train, test in sk.split(data_x, data_y):    x_train = data_x.iloc[train]    y_train = data_y.iloc[train]    x_test = data_x.iloc[test]    y_test = data_y.iloc[test]    # 基准分类模型    #clf = model.fit(x_train,y_train)#, eval_set=(x_test,y_test),cat_features=col)    clf = imbens.ensemble.SelfPacedEnsembleClassifier(        base_estimator = model,#基准分类模型 可以自定义，需要模型包含fit等方法，具体信息请查看参考文献3        n_estimators = 5,        random_state=49    ).fit(x_train, y_train)    yy_pred_valid=clf.predict_proba(x_test)[:,-1]    print('cat验证的auc:{}'.format(roc_auc_score(y_test, yy_pred_valid)))    mean_score += roc_auc_score(y_test, yy_pred_valid) / n_folds    y_pred_valid = clf.predict_proba(testA)[:,-1]    answers.append(y_pred_valid)print('mean valAuc:{}'.format(mean_score))cat_pre=sum(answers)/(n_folds) # 数据比1：4sub['isDefault']=cat_presub.to_csv('金融预测.csv',index=False)

登录后复制    

总结

本项目主要以CBT算法为基分类器对比了几种常用的数据不平衡的处理效果，关于数据特征构建的知识可以查看开头参考文献。

数据的对比提交结果如下：

imbalanced-ensemble库中包含了许多其他的非平衡算法，有兴趣的同学可以关注参考文献[3]

此外，在训练中，随着基模型数量的增加此数据上的分类表现有所下降，这可能是由于此数据失衡仅4：1，不足以体现算法优势。

相关攻略

热门推荐