如何找到数据帧最大 window 的源索引?

How to find source indexes of window max for dataframe?

我有一个 DatetimeIndex 的数据框,我想为每个 window 找到最大元素。但我也必须知道元素的索引。 示例数据:

data = pd.DataFrame(
    index=pd.date_range(start=pd.to_datetime('2010-10-10 12:00:00'),
                        periods=10, freq='H'),
    data={'value': [3, 2, 1, 0, 5, 1, 1, 1, 1, 1]}
)

如果我使用 max 滚动,我会丢失索引:

data.rolling(3).max()

输出:

                     value
2010-10-10 12:00:00    NaN
2010-10-10 13:00:00    NaN
2010-10-10 14:00:00    3.0
2010-10-10 15:00:00    2.0
2010-10-10 16:00:00    5.0
2010-10-10 17:00:00    5.0
2010-10-10 18:00:00    5.0
2010-10-10 19:00:00    1.0
2010-10-10 20:00:00    1.0
2010-10-10 21:00:00    1.0

如果我尝试使用 argmax,我会在每个 window 中获取索引作为整数索引(但我必须找到源日期时间索引或源数据帧的整数索引才能找到它们 iloc):

data.rolling(3).apply(lambda x: x.argmax())

输出:

                     value
2010-10-10 12:00:00    NaN
2010-10-10 13:00:00    NaN
2010-10-10 14:00:00    0.0
2010-10-10 15:00:00    0.0
2010-10-10 16:00:00    2.0
2010-10-10 17:00:00    1.0
2010-10-10 18:00:00    0.0
2010-10-10 19:00:00    0.0
2010-10-10 20:00:00    0.0
2010-10-10 21:00:00    0.0

有谁能帮我在 pandas 中找到好的 function/parameters 吗?

当然我可以使用 for 比如:

pd.DataFrame([{'value_max': data[ind: ind + window][target_var].max(),
               'source_index': data[ind: ind + window].index[data[ind: ind + window][target_var].values.argmax()]
              } for ind in range(1, len(data) + 1 - window)],
             index=data.index[1:-window+1])

并且有效。但我想尝试使用 pandas.

找到更优雅的解决方案

期望输出:

                           source_index  value_max
2010-10-10 13:00:00 2010-10-10 13:00:00          2
2010-10-10 14:00:00 2010-10-10 16:00:00          5
2010-10-10 15:00:00 2010-10-10 16:00:00          5
2010-10-10 16:00:00 2010-10-10 16:00:00          5
2010-10-10 17:00:00 2010-10-10 17:00:00          1
2010-10-10 18:00:00 2010-10-10 18:00:00          1
2010-10-10 19:00:00 2010-10-10 19:00:00          1

尚未为 resampler 实现使用 Resampler.agg with custom function, because idxmax

def idx(x):
    return x.index.values[np.argmax(x.values)]

df = data['value'].rolling(3).agg(['max', idx])
df['idx'] = pd.to_datetime(df['idx'])
print (df)
                     max                 idx
2010-10-10 12:00:00  NaN                 NaT
2010-10-10 13:00:00  NaN                 NaT
2010-10-10 14:00:00  3.0 2010-10-10 12:00:00
2010-10-10 15:00:00  2.0 2010-10-10 13:00:00
2010-10-10 16:00:00  5.0 2010-10-10 16:00:00
2010-10-10 17:00:00  5.0 2010-10-10 16:00:00
2010-10-10 18:00:00  5.0 2010-10-10 16:00:00
2010-10-10 19:00:00  1.0 2010-10-10 17:00:00
2010-10-10 20:00:00  1.0 2010-10-10 18:00:00
2010-10-10 21:00:00  1.0 2010-10-10 19:00:00

谢谢@Sandeep Kadapa 改进解决方案:

def idx(x):
    return x.idxmax().to_datetime64()