本文补充第二种特征生成方法，第一种是A属性等于1，针对一个卡号的B属性各个值求和，还有种方法是统计各个维度上的统计指标（求和等指标）

# In[86]:

transaction=pd.read_csv('d:/transaction_d_pre.csv',low_memory=False)

# In[87]:

transaction.head(5)

# In[98]:

numeric_cols =  ['authorized_flag',  'category_1', 'installments', 'category_3',  'month_lag','purchase_month',
 'purchase_amount', 'category_2', 
 'purchase_month', 'purchase_hour_section',
  'most_recent_sales_range', 'most_recent_purchases_range']

# In[99]:

categorical_cols =  ['city_id', 'merchant_category_id', 'merchant_id', 'state_id', 'subsector_id', 'category_4','purchase_date']

# In[100]:

aggs={}

# In[101]:

for col in numeric_cols:
    aggs[col]= ['nunique', 'mean', 'min', 'max','var','skew', 'sum']

# In[102]:

for col in categorical_cols:
    aggs[col]=['nunique']
aggs

# In[103]:

aggs['card_id']=['size','count']#加上card_id

# In[104]:

cols=['card_id']

# In[105]:

for key in aggs.keys():
    cols.extend([key+''+stat for stat in aggs[key]])#把名称组合出来

     df = transaction[transaction['month_lag']<0].groupby('card_id').agg(aggs).reset_index() 
     df.columns = cols[:1] + [co+'_hist' for co in cols[1:]] 
      
     df2 = transaction[transaction['month_lag']>=0].groupby('card_id').agg(aggs).reset_index() 
     df2.columns = cols[:1] + [co+'_new' for co in cols[1:]] 
     df = pd.merge(df, df2, how='left',on='card_id') 
      
     df2 = transaction.groupby('card_id').agg(aggs).reset_index() 
     df2.columns = cols 
     df = pd.merge(df, df2, how='left',on='card_id') 
     del transaction 
     gc.collect() 

     # 生成训练集与测试集 
     train = pd.merge(train, df, how='left', on='card_id') 
     test =  pd.merge(test, df, how='left', on='card_id') 
     del df 
     train.to_csv(\ preprocess/train_groupby.csv\ , index=False) 
     test.to_csv(\ preprocess/test_groupby.csv\ , index=False) 
      
     gc.collect()