首页 文章

使用Featuretools中的多个训练窗口计算相同的特征

提问于
浏览
2

Featuretools支持已经处理多个截止时间https://docs.featuretools.com/automated_feature_engineering/handling_time.html

In [20]: temporal_cutoffs = ft.make_temporal_cutoffs(cutoffs['customer_id'],
   ....:                                             cutoffs['cutoff_time'],
   ....:                                             window_size='3d',
   ....:                                             num_windows=2)
   ....: 

In [21]: temporal_cutoffs
Out[21]: 
        time  instance_id
0 2011-12-12        13458
1 2011-12-15        13458
2 2012-10-02        13602
3 2012-10-05        13602
4 2012-01-22        15222
5 2012-01-25        15222

In [22]: entityset = ft.demo.load_retail()

In [23]: feature_tensor, feature_defs = ft.dfs(entityset=entityset,
   ....:                                       target_entity='customers',
   ....:                                       cutoff_time=temporal_cutoffs,
   ....:                                       cutoff_time_in_index=True,
   ....:                                       max_features=4)
   ....: 

In [24]: feature_tensor
Out[24]: 
                        MAX(order_products.total)  MIN(order_products.unit_price)  STD(order_products.quantity)  COUNT(order_products)
customer_id time                                                                                                                      
13458.0     2011-12-12                    201.960                          0.3135                     10.053804                    394
            2011-12-15                    201.960                          0.3135                     10.053804                    394
15222.0     2012-01-22                    272.250                          1.1880                     26.832816                      5
            2012-01-25                    272.250                          1.1880                     26.832816                      5
13602.0     2012-10-02                     49.896                          1.0395                      8.732068                     23
            2012-10-05                     49.896                          1.0395                      8.732068                     23

但是,当您在多个时间点看到一个ID时,会生成一个pandas多索引 . 如何(可能通过数据?)我可以改为获得所有MIN / MAX / ...生成的前缀为last_x_days_MIN / MAX / ...的列,以便每个截止窗口获得额外的功能?

编辑所需的输出格式

initial feature 1,initial feature 2, time_frame_1_<AGGTYPE2>_Feature,time_frame_1_<AGGTYPE1>_Feature,time_frame_2_<AGGTYPE1>_Feature,time_frame_2_<AGGTYPE2>_Feature,time_frame_2_<AGGTYPE1>_Feature,time_frame_2_<AGGTYPE1>_Feature

1 回答

  • 3

    您可以通过使用不同的 training_windowsft.calculate_feature_matrix 进行两次调用并将生成的要素矩阵连接在一起来实现此目的 . 例如,

    import featuretools as ft
    import pandas as pd
    
    entityset = ft.demo.load_retail()
    
    cutoffs = pd.DataFrame({
          'customer_name': ["Micheal Nicholson", "Krista Maddox"],
          'cutoff_time': [pd.Timestamp('2011-10-14'), pd.Timestamp('2011-08-18')]
        })
    
    feature_defs = ft.dfs(entityset=entityset,
                          target_entity='customers',
                          agg_primitives=["sum"],
                          trans_primitives=[],
                          max_features=1,
                          features_only=True)
    
    
    
    fm_60_days = ft.calculate_feature_matrix(entityset=entityset,
                                             features=feature_defs,
                                             cutoff_time=cutoffs,
                                             training_window="60 days")
    
    fm_30_days = ft.calculate_feature_matrix(entityset=entityset,
                                             features=feature_defs,
                                             cutoff_time=cutoffs,
                                             training_window="30 days")
    
    fm_60_days.merge(fm_30_days, left_index=True, right_index=True, suffixes=("__60_days", "__30_days"))
    

    上面的代码返回此DataFrame,其中我们使用最后60天和30天的数据计算相同的功能 .

    SUM(order_products.quantity)__60_days  SUM(order_products.quantity)__30_days
    customer_name                                                                                  
    Krista Maddox                                        466                                    306
    Micheal Nicholson                                    710                                    539
    

    注意:此示例在最新版本的Featuretools(v0.3.1)上运行,其中我们更新了演示零售数据集,以将可解释的名称作为客户ID .

相关问题