便签：Xgboost
[TOC]
背景：对于投资者来说，需要预测借贷者无法偿还贷款的风险，最大程度地避免投资损失，最大程度地实现投资回报。如果借贷者顺利偿还贷款，投资者则获得利息收益。如果借贷者无法偿还贷款，投资者则损失贷款本金
目标：根据贷款记录数据构建机器学习模型，预测借贷者是否有信用问题
-
该项目的预测问题为二分类问题，预测用户是否是“坏人”
-
特征未进行脱敏处理，可以利用业务知识对特征进行先处理后再进行常规特征工程
-
在金融领域，容易出现较严重的正负样本不平衡问题，且特征容易出现共线性的情况，因而使用GBRT作为本次问题的模型
df_clean.loan_status.replace('Fully Paid', 0, inplace = True)
df_clean.loan_status.replace('Current', 0, inplace = True)
df_clean.loan_status.replace('Late (16-30 days)', 1, inplace = True)
df_clean.loan_status.replace('Late (31-120 days)', 1, inplace = True)
df_clean.loan_status.replace('Charged Off', 1, inplace = True)
df_clean.loan_status.replace('In Grace Period', np.nan, inplace = True)
df_clean.loan_status.replace('Default', np.nan, inplace = True)
df_clean.loan_status.value_counts(dropna = False)
df_clean.dropna(subset = ['loan_status'], inplace = True)
y = df_clean.loan_status
df_clean.drop('loan_status', axis = 1, inplace = True)- 把已经还款和正在还款的客户定义为”好“客户，把逾期和已有征信危机的客户定义为“坏”客户
df_clean = df.copy()
df_clean.drop('member_id', axis = 1, inplace = True)
df_clean.drop('id', axis = 1, inplace = True)
df_clean.dropna(axis = 0, how = 'all', inplace = True)
df_clean.dropna(axis = 1, how = 'all', inplace = True)missing_col = []
missing_col_nums = []
for col in df_clean.columns:
missing_nums = round(len(df_clean[df_clean[col].isnull()].index) / len(df_clean), 5)
if missing_nums:
missing_col.append(col)
missing_col_nums.append((col, missing_nums))
for i in sorted(missing_col_nums, key = lambda x: x[1], reverse = True):
print(i)
if i[1] >0.5:
missing_col.remove(i[0])
missing_col_nums.remove(i)
df_clean.drop(i[0], axis = 1, inplace = True)df_clean.emp_length.fillna(0, inplace = True)
df_clean.emp_length.replace(to_replace = '[^0-9]+', value = '', inplace = True, regex = True)
df_clean.emp_length = df_clean.emp_length.astype(int)
#将百分比数据处理成int型
df_clean.revol_util = df_clean.revol_util.str.replace('%', "").astype(float)
df_clean.int_rate = df_clean.int_rate.str.replace('%','').astype(float)
df_clean.drop(['out_prncp','out_prncp_inv','total_pymnt',
'total_pymnt_inv','total_rec_prncp', 'grade', 'sub_grade'] ,1, inplace=True)
df_clean.drop(['total_rec_int','total_rec_late_fee',
'recoveries','collection_recovery_fee',
'collection_recovery_fee' ],1, inplace=True)
df_clean.drop(['last_pymnt_d','last_pymnt_amnt',
'next_pymnt_d','last_credit_pull_d'],1, inplace=True)
df_clean.drop(['policy_code'],1, inplace=True)df_clean.select_dtypes(include = ['float']).describe().T.sort_values(by = 'std')
float_col = list(df_clean.select_dtypes(include = ['float']).columns)
for col in float_col:
if df_clean[col].std() <= 1.5:
df_clean.drop(col, axis = 1, inplace = True)corr_target = []
for col in df_clean.select_dtypes(exclude = ['object']).columns:
corr = np.corrcoef(df_clean[col].fillna(df_clean[col].median()), y)[0][1]
corr_target.append((col, corr))
for i in sorted(corr_target, key = lambda x: abs(x[1]), reverse = False):
print(i)df_clean['label'] = y
plt.figure(figsize = (32, 24))
sns.heatmap(df_clean.corr().round(5), annot = True, cmap = 'RdBu')
df_clean.drop('label', axis = 1, inplace = True)
- 通过图像暂未发现有完全不相关的特征
cor = df_clean.corr()
cor.loc[ : , : ] = np.tril(df_clean.corr().round(5), k = -1)
cor = cor.stack()
cor_set = set()
for index in cor[(cor > 0.8)| (cor < -0.8].index:
cor_set.add(index[0])
for col in cor_set:
df_clean.drop(col, axis = 1, inplace = True)
plt.figure(figsize = (32, 24))
sns.heatmap(df_clean.select_dtypes(exclude = ['object']).corr().round(2), /
annot = True, cmap = 'RdBu')
k = 0
plt.figure(figsize = (32, 24))
for col in df_clean.select_dtypes(exclude = ['object']).columns:
k += 1
plt.subplot(7, 10, k)
plt.ylabel('Counts')
sns.distplot(df_clean[col].dropna(), kde = False)
- 数值型数据分布不是正态分布，后续填充空值时使用中值进行填充，保证分布不被改变
object_list = []
object_col = []
for col in df_clean.select_dtypes(include = 'object').columns:
object_list.append((col, len(df_clean[col].unique())))
object_col.append(col)
for i in sorted(object_list, key = lambda x: x[1], reverse = True):
print(i)
if i[1] >= 50:
object_list.remove(i)
object_col.remove(i[0])
df_clean.drop(i[0], axis = 1, inplace = True)### 4.10、总体缺失数据的相关性分析df_clean_object = df_clean.select_dtypes(include = ['object']).copy()
df_clean_object.fillna('missing', inplace = True)
k = 0
plt.figure(figsize = (64, 12))
for col in df_clean_object.columns:
val = df_clean_object[col].value_counts()
k += 1
plt.subplot(4, 3, k)
names = list(val.index)
counts = list(val.values)
data = pd.DataFrame({'names': names, 'counts': counts})
plt.title(col)
sns.barplot(x = names, y = counts, data = data)
lean.drop(['application_type', 'pymnt_plan'], axis = 1, inplace = True)
- mo_sin_old_il_acct与mths_since_rcnt_il的缺失情况完全正相关，后续需要为其构建特征
- mths_since_recent_bc与bc_open_to_buy的缺失情况正相关，后续需要为其构建特征
- 根据数据的缺失情况构建新特征
df_clean['col_miss_1'] = np.zeros(df_clean.shape[0])
df_clean['col_miss_2'] = np.zeros(df_clean.shape[1])
df_clean.loc[df_clean['mo_sin_old_il_acct'].isnull(), 'col_miss_1'] += 1
df_clean.loc[df_clean['mths_since_rcnt_il'].isnull(), 'col_miss_1'] += 1
df_clean.loc[df_clean['mths_since_recent_bc'].isnull(), 'col_miss_2'] += 1
df_clean.loc[df_clean['bc_open_to_buy'].isnull(), 'col_miss_2'] += 1df_clean = pd.get_dummies(df_clean)from sklearn.impute import SimpleImputer as imputer
impute = imputer(strategy = 'median')
df_clean.iloc[ :, : ] = impute.fit_transform(df_clean)from sklearn.model_selection import train_test_split as tts
X_train, X_test, y_train, y_test = tts(df_clean, y, test_size = 0.2,
random_state = 2019, stratify = y)import xgboost as xgb
dtrain = xgb.DMatrix(X_train, y_train)
dtest = xgb.DMatrix(X_test, y_test)
param = {'verbosity': 1,
'objective': 'binary:logistic',
'stratified': True,
'max_depth': 1,
'subsample': 0.8,
'eta': 0.1,
'gamma': 0,
'lambda': 4,
'alpha': 0,
'scale_pos_weight': 20,
'min_child_weight': 5,
'max_delta_step': 0.5,
'base_score': 0.05,
'colsample_bytree': 0.6,
'colsample_bynode': 0.4,
'colsample_bylevel': 0.4,
'seed': 2019,
'eval_metric': 'auc',
'nfold': 4}
num_round = 1000
pred_cv = xgb.cv(param, dtrain, num_round, nfold = 4, verbose_eval = False)
print(pred_cv.tail())
plt.figure(figsize = (12, 6))
sns.lineplot(data = pred_cv.drop(['train-error-std', 'test-error-std'], axis = 1))
将num_round设为150
dtrain = xgb.DMatrix(X_train, y_train)
dtest = xgb.DMatrix(X_test, y_test)
param = {'verbosity': 1,
'objective': 'binary:logistic',
'stratified': True,
'max_depth': 1,
'subsample': 0.8,
'eta': 0.1,
'gamma': 0,
'lambda': 4,
'alpha': 0,
'scale_pos_weight': 20,
'min_child_weight': 5,
'max_delta_step': 0.5,
'base_score': 0.05,
'colsample_bytree': 0.6,
'colsample_bynode': 0.4,
'colsample_bylevel': 0.4,
'seed': 2019,
'eval_metric': 'auc',
'nfold': 4}
num_round = 150
bst = xgb.train(param, dtrain, num_round)
pred = bst.predict(dtest)- 使用scale_pos_weight和base_score可以减少迭代次数
y_pred = pred.copy()
y_pred[y_pred > 0.5] = 1
y_pred[y_pred != 1] = 0
print('AUC is ', auc(y_test, y_pred))
print('Recall_rate is ', recall(y_test, y_pred))
print('Logloss is', log_loss(y_test, pred))
fpr, tpr, threadshold = roc_curve(y_test, pred)
ks = max(tpr - fpr)
print('KS系数为', ks)
plt.figure(figsize = (16,12))
sns.lineplot(data = pd.DataFrame({'fpr': fpr, 'tpr': tpr, 'max_ks': tpr - fpr},
index = threadshold))
项目结果：KS系数为0.42