【数据竞赛】风控实操案例 | 基于Xgboost与Catboost实现非法集资企业识别

import osos.chdir('/home/datafountain/work/')

# !unzip pictures.zip

# 使用pip命令安装指定版本工具包# !pip install jieba tqdm scikit-learn catboost!pip install catboost -i  https://pypi.doubanio.com/simple/

# !pip install toad==0.0.61

#统一导入工具包import pandas as pdimport numpy as npfrom sklearn.model_selection import StratifiedKFoldfrom catboost import CatBoostClassifierfrom sklearn.preprocessing import LabelEncoderimport gcimport osfrom sklearn.metrics import precision_score, recall_score, f1_scoreimport xgboost as xgbimport warningswarnings.filterwarnings("ignore")import jiebaimport matplotlib.pyplot as plt%matplotlib inline

# Debug参数  DEBUG = True  # Debug模式可快速跑通代码，非Debug模式可得到更好的结果
Iterations=20 if DEBUG else 5000

!mkdir datasets/  submit/  temp/!mkdir datasets/data  datasets/pretrained!unzip -o ../datasets/test.zip -d datasets/data!unzip -o ../datasets/train.zip -d datasets/datacp ../datasets/cn_stopwords.txt  /home/datafountain/work/datasets/pretrained/

# 读取数据base = pd.read_csv('./datasets/data/train/base_info.csv')label = pd.read_csv('./datasets/data/train/entprise_info.csv')entprise_evaluate = pd.read_csv('./datasets/data/test/entprise_evaluate.csv')
train = label.merge(base, on='id', how='left') test = entprise_evaluate.merge(base, on='id', how='left')target = train.label
target.value_counts().plot.bar()

print(target.value_counts(normalize=True))

import toadtoad.detect(train)

toad.detect(test)

# 删除缺失率较高的列base = base.drop(columns=['congro',                          'exenum',                          'forreccap',                          'forregcap',                          'midpreindcode',                          'parnum',                          'protype',                          'ptbusscope'])

# 目标编码函数def kfold_stats_feature(train, test, feats, k, seed):    '''    Target-Encode    '''    folds = StratifiedKFold(n_splits=k, shuffle=True, random_state=seed)  # 这里最好和后面模型的K折交叉验证保持一致…… ……    del train['fold']    return train, test
# 初始化种子与交叉验证折数seed = 2020k = 5…… ……del datagc.collect()print('done...')

#  可解释性特征def gen_annual(annual_report):    '''    年报表特征提取    '''    新闻表特征提取    '''    变更表特征提取    '''    报税表特征提取    '''

# tfidif 处理经营范围的特征    # 创建一个停用词列表    stopwords = [line.strip() for line in open('./datasets/pretrained/cn_stopwords.txt', encoding='UTF-8').readlines()]            # 结巴分词        ……         ……  print('对opscope提取tfidif特征完毕..........')    base_info[['id', 'tfidif_opscope']].to_csv('./temp/tfidf_op.csv', index=False)gen_tfidf()

# opscope 字段中是否包含投资train['opscope_info'] = train.opscope.apply(lambda x: 1 if '投资' in x else 0)test['opscope_info'] = test.opscope.apply(lambda x: 1 if '投资' in x else 0)

def cal_f1(train_y, train_pred, test_y, val_pred):    train_pred[train_pred >= 0.5] = 1    train_pred[train_pred < 0.5] = 0    ##train … … # 导入数据    print('读取数据...')    ……    sub_df.to_csv(        './temp/cat_{}.csv'.format(model_num),        index=False)    return sub_df
cat1 = cat_model(False, 1)cat2 = cat_model(True, 2)

def f1_score_eval(preds, valid_df):    '''    自定义XGB的早停函数    '''     # ***导入数据*** #    print('读取数据...')    base = pd.read_csv('./datasets/data/train/base_info.csv')    '''print('OOF-MEAN-AUC:%.6f, OOF-STD-AUC:%.6f' % (np.mean(offline_score), np.std(offline_score))) print(feature_importance_df.groupby(['Feature'])['importance'].mean().sort_values(ascending=False)) # 特征重要性 sub_df = test[['id']].copy().reset_index(drop=True) sub_df['score'] = output_preds sub_df.to_csv(        './temp/xgb_{}.csv'.format(model_num), index=False) return sub_df    
xgb1 = xgb_model(False, 1)xgb2 = xgb_model(True, 2)

def voting(ll):    for i, oof in enumerate(ll):        oof[f'score_{i}'] = oof.score.apply(lambda x: 1 if x > 0.5 else 0)
    sub = ll[0].id.copy()    sub = pd.DataFrame(sub)    for i, tmp in enumerate(ll):        sub = sub.merge(tmp, on='id', how='left')    sub['score'] = sub.iloc[:, 1:].apply(lambda x:x.mode()[0], axis=1)    sub[['id', 'score']].to_csv('./submit/result.csv', index=False)    return sub[['id', 'score']]

fusion.score.value_counts().plot.bar()

fusion.score.value_counts()