首页 > 解决方案 > (Featuretools) 如何计算聚合特征图元?

问题描述

即使我使用非常简单的数据进行测试,我也不知道如何计算聚合特征原语。我也查看了 featuretools 代码,但找不到聚合操作发生在哪里。

以下是示例代码:

from sklearn.utils import shuffle

periods = 5
end_date = "2012-04-13"
train_df = pd.DataFrame(
    {
        "store_id": [0]*periods + [1]*periods + [2]*periods + [3]*periods,
        "region": ["A"]*periods+["B"]*periods*3,
        "amount": shuffle(range(periods*4)),
        "transacted_date": [
            "2012-02-05", "2012-02-10", "2012-03-01", "2012-03-18", "2012-04-23",
        ]*4
    }
)
train_df["transacted_date"] = pd.to_datetime(train_df["transacted_date"])
train_df.sort_values(["store_id", "transacted_date"], inplace=True)


def make_retail_cutoffs_amounts(data_df, amount_start_date, amount_end_date):
    store_pool = data_df[data_df['transacted_date'] < amount_start_date]['store_id'].unique()
    tmp = pd.DataFrame({'store_id': store_pool})

    amounts = data_df[
        (data_df['store_id'].isin(store_pool)) &
        (amount_start_date <= data_df['transacted_date']) &
        (data_df['transacted_date'] < amount_end_date)
    ].groupby('store_id')['amount'].sum().reset_index()

    amounts = amounts.merge(tmp, on = 'store_id', how = 'right')
    amounts['amount'] = amounts['amount'].fillna(0)  # 0으로 채워지는 애는 3개월 다 수익이 없는 녀석!

    amounts['cutoff_time'] = pd.to_datetime(amount_start_date)

    amounts = amounts[['store_id', 'cutoff_time', 'amount']]
    amounts = amounts.rename(columns={"amount":"1month_amount_from_cutoff_time"})
    return amounts


amount_start_date = "2012-02-01"
amount_end_date = end_date
agg_month = 1

data_df_list = []
date_list = pd.date_range(amount_start_date, datetime.strptime(end_date, "%Y-%m-%d") + pd.DateOffset(months=1), freq="MS")

for amount_start_date, amount_end_date in zip(date_list[:-agg_month], date_list[agg_month:]):
    data_df_list.append(
        make_retail_cutoffs_amounts(
            train_df, amount_start_date, amount_end_date
        )
    )
data_df = pd.concat(data_df_list)
data_df.sort_values(["store_id", "cutoff_time", ], inplace=True)

import featuretools as ft

es = ft.EntitySet(id="sale_set")
es = es.entity_from_dataframe(
    "sales",
    dataframe=train_df,
    index="sale_id", make_index=True,
    time_index='transacted_date',
)
es.normalize_entity(
    new_entity_id="stores",
    base_entity_id="sales",
    index="store_id",
    additional_variables=['region']
)

# When using a training window, 
# it is necessary to calculate the last time indexes for the entity set. Adding
es.add_last_time_indexes()

features  = ft.dfs(
    entityset=es,
    target_entity='stores',
    cutoff_time=data_df,
    verbose=1,
    cutoff_time_in_index=True,
    n_jobs=1,
    max_depth=2,

    agg_primitives=["sum",],
    trans_primitives=["cum_max"], 
    training_window="1 month",
)

dfs工作正常,但无法解释结果特征。

这是特征的样本数据:

在此处输入图像描述

正如您在此处看到的,第一行SUM(sales.amount)SUM(sales.CUM_MAX(amount))分别是 19 和 37。我想知道它们是如何计算的。

以下是我对结果的解释:

在此处输入图像描述

  1. 如您在此处看到的,store_0 在 2012 年 2 月有 2 条销售数据记录。因此,SUM(sales.amount)截止时间 2012-03-01 的 store_id=0 应该是 0 + 8 = 8,而不是 19。

  2. 同样,SUM(sales.CUM_MAX(amount))截止时间为 2012-03-01 的 store_id=0 也应该是 SUM(sales.CUM_MAX(amount)) = SUM([0, 8]) = 8,而不是 37。

我错过了什么吗?它们是如何计算的?

标签: pythonfeaturetools

解决方案


这些概念将帮助您了解如何计算特征:

  1. 时间索引达到并包括截止时间的数据用于计算特征。
  2. 为了CumMax在每个商店的基础上进行计算,它需要放置在groupby_trans_primitives而不是trans_primitives.

我将通过一个例子。这是 CSV 格式的数据。

store_id,region,amount,transacted_date
0,A,16,2012-02-05
0,A,15,2012-02-10
0,A,13,2012-03-01
0,A,2,2012-03-18
0,A,3,2012-04-23
1,B,9,2012-02-05
1,B,8,2012-02-10
1,B,14,2012-03-01
1,B,1,2012-03-18
1,B,5,2012-04-23
2,B,6,2012-02-05
2,B,12,2012-02-10
2,B,4,2012-03-01
2,B,7,2012-03-18
2,B,11,2012-04-23
3,B,18,2012-02-05
3,B,19,2012-02-10
3,B,0,2012-03-01
3,B,10,2012-03-18
3,B,17,2012-04-23

首先,我加载数据集。

import pandas as pd

train_df = pd.read_csv('data.csv', parse_dates=['transacted_date'])
train_df.sort_values(["store_id", "transacted_date"], inplace=True)
 store_id region  amount transacted_date
        0      A      16      2012-02-05
        0      A      15      2012-02-10
        0      A      13      2012-03-01
        0      A       2      2012-03-18
        0      A       3      2012-04-23
        1      B       9      2012-02-05
        1      B       8      2012-02-10
        1      B      14      2012-03-01
        1      B       1      2012-03-18
        1      B       5      2012-04-23
        2      B       6      2012-02-05
        2      B      12      2012-02-10
        2      B       4      2012-03-01
        2      B       7      2012-03-18
        2      B      11      2012-04-23
        3      B      18      2012-02-05
        3      B      19      2012-02-10
        3      B       0      2012-03-01
        3      B      10      2012-03-18
        3      B      17      2012-04-23

然后,我使用Compose自动生成相同的截止时间。

import composeml as cp

def total_amount(df):
    return df.amount.sum()


lm = cp.LabelMaker(
    target_entity='store_id',
    time_index='transacted_date',
    labeling_function=total_amount,
    window_size='1MS',
)

lt = lm.search(
    train_df,
    num_examples_per_instance=-1,
    minimum_data='2012-03-01',
)
 store_id cutoff_time  total_amount
        0  2012-03-01            15
        0  2012-04-01             3
        1  2012-03-01            15
        1  2012-04-01             5
        2  2012-03-01            11
        2  2012-04-01            11
        3  2012-03-01            10
        3  2012-04-01            17

现在,我构建实体集。

import featuretools as ft

es = ft.EntitySet(id="sale_set")

es = es.entity_from_dataframe(
    "sales",
    dataframe=train_df,
    index="sale_id",
    make_index=True,
    time_index='transacted_date',
)

es.normalize_entity(
    new_entity_id="stores",
    base_entity_id="sales",
    index="store_id",
    additional_variables=['region'],
)

es.add_last_time_indexes()

这是一个了解实体集结构的图。

es.plot()

在此处输入图像描述

最后,我运行 DFS 来计算特征。

fm, fd = ft.dfs(
    entityset=es,
    target_entity='stores',
    cutoff_time=lt,
    cutoff_time_in_index=True,
    agg_primitives=["sum"],
    groupby_trans_primitives=["cum_max"],
    training_window="1 month",
    max_depth=2,
    verbose=1,
)

fm.filter(regex='SUM')
                     SUM(sales.amount)  SUM(sales.CUM_MAX(amount) by store_id)
store_id time
0        2012-03-01                 44                                    48.0
         2012-04-01                 15                                    26.0
1        2012-03-01                 31                                    32.0
         2012-04-01                 15                                    28.0
2        2012-03-01                 22                                    30.0
         2012-04-01                 11                                    11.0
3        2012-03-01                 37                                    56.0
         2012-04-01                 10                                    10.0

我们可以看到SUM(sales.amount)for store 0at cutoff time2012-03-01包括直到并包括2012-03-01计算特征的数据。

>>> 16 + 15 + 13
44

在 DFS 中SUM(sales.CUM_MAX(amount) by store_id)作为 a 应用时也是如此。groupby_trans_primitives

>>> sum((16, 16, 16))
48

让我知道这是否有帮助。


推荐阅读