本例程是通过客户提供的信息分析客户会产生违约的可能性。由此来判断是否要给客户提供贷款。背景内容不再多说，数据相关的解释在代码中会有注释。运行中缺失的包请自行安装，我这里的环境是anaconda
直接上代码：
import numpy as noimport pandas as pdimport osimport seaborn as snscolor = sns.color_palette()import matplotlib.pyplot as plt%matplotlib inlineimport plotly.offline as pypy.init_notebook_mode(connected=True)from plotly.offline import init_notebook_mode,iplotinit_notebook_mode(connected=True)import plotly.graph_objs as goimport plotly.offline as offlineoffline.init_notebook_mode()import cufflinks as cfcf.go_offline()
#下面开始加载数据df_train = pd.read_csv('./dataset/Home_Credit/application_train.csv')df_test = pd.read_csv('./dataset/Home_Credit/application_test.csv')
#看看都有哪些属性df_train.columns.values#属性很多，有点吓人
print(df_train.shape)#(307511, 122)df_train.head()
#检查application_train 中的缺失数据total = df_train.isnull().sum().sort_values(ascending = False)percent = (df_train.isnull().sum()/df_train.isnull().count()*100).sort_values(ascending=False)missing_application_train_data = pd.concat([total,percent],axis = 1,keys=['Toatl','Percent'])missing_application_train_data.head(10)
#开始探索我们的数据#贷款金额 分布plt.figure(figsize=(12,5))plt.title("Distribution of AMT_CREDIT")ax = sns.distplot(df_train["AMT_CREDIT"])
#客户年收入(大部分人都是在50000以下)plt.figure(figsize=(12,5))plt.title("Distribution of AMT_INCOME_TOTAL")ax = sns.distplot(df_train["AMT_ANNUITY"].dropna())
#消费贷款，对应贷款的商品的价格plt.figure(figsize=(12,5))plt.title("Distribution of AMT_GOODS_PRICE")ax = sns.distplot(df_train['AMT_GOODS_PRICE'].dropna())
#申请贷款的时候客户的陪同人temp = df_train["NAME_TYPE_SUITE"].value_counts()trace = go.Bar(x = temp.index,y = (temp / temp.sum())*100,)data = [trace]layout = go.Layout(title = "Distribution of Name of type of the Suite in % ",xaxis=dict(title='Name of type of the Suite',tickfont=dict(size=14,color='rgb(107, 107, 107)')),yaxis=dict(title='Count of Name of type of the Suite in %',titlefont=dict(size=16,color='rgb(107, 107, 107)'),tickfont=dict(size=14,color='rgb(107, 107, 107)')))fig = go.Figure(data=data, layout=layout)py.iplot(fig, filename='schoolStateNames')#如下图看来一个人的时候发生贷款的概率更高，有人陪同估计不好意思
#是否发生逾期未还的情况分布,看来绝大多数人还是守信用的temp = df_train["TARGET"].value_counts()df = pd.DataFrame({'labels': temp.index,'values': temp.values})df.iplot(kind='pie',labels='labels',values='values', title='Loan Repayed or not')
#贷款是现金还是循环的标识 （就是一次性拿到全部贷款还是当前只拿部分在后面需要的时候再拿）temp = df_train["NAME_CONTRACT_TYPE"].value_counts()fig = {"data": [{"values": temp.values,"labels": temp.index,"domain": {"x": [0, .48]},#"name": "Types of Loans",#"hoverinfo":"label+percent+name","hole": .7,"type": "pie"},],"layout": {"title":"Types of loan","annotations": [{"font": {"size": 20},"showarrow": False,"text": "Loan Types","x": 0.17,"y": 0.5}]}}iplot(fig, filename='donut')#如下图可知绝大部分人都只是会拿到贷款全部额度，毕竟贷款一般是解燃眉之急，很少有人贷款回来慢慢用。
#是否有房/车#FLAG_OWN_CAR 客户是否拥有汽车#FLAG_OWN_REALTY 客户是否拥有房屋或公寓temp1 = df_train["FLAG_OWN_CAR"].value_counts()temp2 = df_train["FLAG_OWN_REALTY"].value_counts()fig = {"data": [{"values": temp1.values,"labels": temp1.index,"domain": {"x": [0, .48]},"name": "Own Car","hoverinfo":"label+percent+name","hole": .6,"type": "pie"},{"values": temp2.values,"labels": temp2.index,"textposition":"inside","domain": {"x": [.52, 1]},"name": "Own Reality","hoverinfo":"label+percent+name","hole": .6,"type": "pie"}],"layout": {"title":"Purpose of loan","annotations": [{"font": {"size": 20},"showarrow": False,"text": "Own Car","x": 0.20,"y": 0.5},{"font": {"size": 20},"showarrow": False,"text": "Own Reality","x": 0.8,"y": 0.5}]}}iplot(fig, filename='donut')#如下图看下来，多数贷款的人是有房没车的人。有房没车估计也是底层人民啊，这符合我们正常的认知，没有住所的人去贷款估计也很难通过（谁愿意借钱给流浪汉呢）
# 收入类型# 工作/商业助理/退休人员/公务员/失业/学生/商人/产假temp = df_train['NAME_INCOME_TYPE'].value_counts()df = pd.DataFrame({'labels':temp.index,'values':temp.values})df.iplot(kind='pie',labels='labels',values='values',title='Income sources of Applicant\'s',hole=0.5)#多数人还是上班族（干得多拿得少，万恶的资本主义）
#贷款申请人的家庭状况#结婚（有宗教或教堂参与的）/单身/民事婚姻（类似中国有政府部门颁发结婚证的民间组织的婚姻）/分离/寡（应该是丧偶）/未知temp = df_train['NAME_FAMILY_STATUS'].value_counts()df = pd.DataFrame({'labels':temp.index,'values':temp.values})df.iplot(kind='pie',labels='labels',values='values',title='Family Status of Applicant\'s',hole=0.6)#除了正常已婚人士，单身汉也不少，看来单身汉是真缺钱（要不然也不会单身是吧）
#申请人的职业temp = df_train['OCCUPATION_TYPE'].value_counts()# df = pd.DataFrame({'labels':temp.index,# 'values':temp.values})# df.iplot(kind='pie',labels='labels',values='values',title='Family Status of Applicant\'s',hole=0.6)temp.iplot(kind='bar',xTitle='Occupation',yTitle='Count',title='Occupation of Applicatnt\'s who applied for loan',color='green')#看看下图，最缺钱的是伟大的劳动者，最不缺钱的竟然是我们IT人员（看来是我拖大家的后腿了）
#申请人的教育情况temp = df_train['NAME_EDUCATION_TYPE'].value_counts()df = pd.DataFrame({'labels':temp.index,'values':temp.values})df.iplot(kind='pie',labels='labels',values='values',title='Education od Applicant\'s',hole=0.5)#Secondary special 中等专业学校学历的人最缺钱，然后是Higher education高等教育，难道是学历越高眼界越高，欲望越多，压力越大（也有可能是其他的情况，比如学历底了收入少、还款能力底，贷款批不下来，也就不再去申请贷款了）
#住房情况temp = df_train["NAME_HOUSING_TYPE"].value_counts()df = pd.DataFrame({'labels': temp.index,'values': temp.values})df.iplot(kind='pie',labels='labels',values='values', title='Type of House', hole = 0.5)#住父母房子的人贷款的是最多的（难道是生活压力小只想这享乐了，合租的人很少去贷款估计是要攒钱改善生活吧）
#工作机构类型temp = df_train["ORGANIZATION_TYPE"].value_counts()df = pd.DataFrame({'labels': temp.index,'values': temp.values})df.iplot(kind='pie',labels='labels',values='values', title='Type of House', hole = 0.5)#最缺钱的是做实体的（这个国内情况很相似，踏实做事的企业赚不到钱；反倒不如投机倒把，炒房，炒股票的赚钱，堪忧啊）
#将类别属性数值化from sklearn import preprocessing#找出类别的属性categorical_features = [categorical for categorical in df_train.columns if df_train[categorical].dtype == 'object']#将类别属性数值化for i in categorical_features:lben = preprocessing.LabelEncoder()lben.fit(list(df_train[i].values.astype('str')) + list(df_test[i].values.astype('str')))df_train[i] = lben.transform(list(df_train[i].values.astype('str')))df_test[i] = lben.transform(list(df_test[i].values.astype('str')))
#用-999填充空值df_train.fillna(-999, inplace = True)df_test.fillna(-999, inplace = True)
#构建模型#LightGBM是个快速的，分布式的，高性能的基于决策树算法的梯度提升框架。可用于排序，分类，回归以及很多其他的机器学习任务中。#如果没有lightgbm包，则需要安装（用了镜像源） pip install -i https://pypi.tuna./simple/ lightgbmimport lightgbm as lgbfrom sklearn.model_selection import train_test_split
#提取标签列Y = df_train['TARGET']test_id = df_test['SK_ID_CURR']#删除不用与训练的属性train_X = df_train.drop(['TARGET','SK_ID_CURR'],axis=1)test_X = df_test.drop(['SK_ID_CURR'], axis = 1)
#训练集分割为训练数据和验证数据x_train, x_val, y_train, y_val = train_test_split(train_X, Y, random_state=18)lgb_train = lgb.Dataset(data=x_train, label=y_train)lgb_eval = lgb.Dataset(data=x_val, label=y_val)
#模型参数params = {'task': 'train', 'boosting_type': 'gbdt', 'objective': 'binary', 'metric': 'auc', 'learning_rate': 0.05, 'num_leaves': 32, 'num_iteration': 500, 'verbose': 0 }
#开始训练model = lgb.train(params,lgb_train,valid_sets=lgb_eval,early_stopping_rounds=100,verbose_eval=10)
#特征的重要性分布如下lgb.plot_importance(model,figsize=(18,20))
#预测pred = model.predict(test_X)sub = pd.DataFrame()sub['SK_ID_CURR'] = test_idsub['TARGET'] = pred#保存结果sub.to_csv("baseline_submission.csv", index=False)sub.head(10)
#换一个训练模型#LGBMClassifierfrom lightgbm import LGBMClassifierclf = LGBMClassifier(n_estimators=300,num_leaves=15,colsample_bytree=.8,subsample=.8,max_depth=7,reg_alpha=.1,reg_lambda=.1,min_split_gain=0.01)
#开始训练clf.fit(x_train,y_train,eval_set=[(x_train,y_train),(x_val,y_val)],eval_metric='auc',verbose=0,early_stopping_rounds=30)
#预测pred_1 = clf.predict(test_X)sub = pd.DataFrame()sub['SK_ID_CURR'] = test_idsub['TARGET'] = pred_1sub.to_csv("submission_clf.csv", index=False)sub.head(10)
以上便是通过客户提供的信息预测客户有可能违约的模型实现过程，这里我将有数据都纳如到训练中，当然也可以根据你你自己的判断和思考去掉某些属性；也可以对其中的一些数值型属性进行分段划分。另外也可以用其他你认为更好的算法来训练模型，欢迎流言交流。
