熊猫：枚举索引中的重复项

Question

假设我有一系列发生在不同按键上的事件。

data = [
    {"key": "A", "event": "created"},
    {"key": "A", "event": "updated"},
    {"key": "A", "event": "updated"},
    {"key": "A", "event": "updated"},
    {"key": "B", "event": "created"},
    {"key": "B", "event": "updated"},
    {"key": "B", "event": "updated"},
    {"key": "C", "event": "created"},
    {"key": "C", "event": "updated"},
    {"key": "C", "event": "updated"},
    {"key": "C", "event": "updated"},
    {"key": "C", "event": "updated"},
    {"key": "C", "event": "updated"},
]

df = pandas.DataFrame(data)

我想先在键上索引我的DataFrame，然后再枚举。看起来像是简单的取消堆栈操作，但是我找不到正确的操作方法。

我能做的最好的是

df.set_index("key", append=True).swaplevel(0, 1)

          event
key            
A   0   created
    1   updated
    2   updated
    3   updated
B   4   created
    5   updated
    6   updated
C   7   created
    8   updated
    9   updated
    10  updated
    11  updated
    12  updated

但是我期望的是

          event
key            
A   0   created
    1   updated
    2   updated
    3   updated
B   0   created
    1   updated
    2   updated
C   0   created
    1   updated
    2   updated
    3   updated
    4   updated
    5   updated

我也尝试过

df.groupby("key")["key"].count().apply(range).apply(pandas.Series).stack()

，但是不保留顺序，因此无法将结果用作索引。此外，对于看起来很标准的手术，我觉得有些矫kill过正...

有什么主意吗？

Answer 1

groupby + cumcount

以下是几种方法：

# new version thanks @ScottBoston
df = df.set_index(['key', df.groupby('key').cumcount()])\
       .rename_axis(['key','count'])

# original version
df = df.assign(count=df.groupby('key').cumcount())\
       .set_index(['key', 'count'])

print(df)

             event
key count         
A   0      created
    1      updated
    2      updated
    3      updated
B   0      created
    1      updated
    2      updated
C   0      created
    1      updated
    2      updated
    3      updated
    4      updated
    5      updated

Answer 2

您可以在numpy中执行以下操作：

# df like in OP
keys = df['key'].values
# detect indices where key changes value
change = np.zeros(keys.size, dtype=int)
change[1:] = keys[1:] != keys[:-1]
# naive sequential number
seq = np.arange(keys.size)
# offset by seq at most recent change
offset = np.maximum.accumulate(change * seq)
df['seq'] = seq - offset
print(df.set_index(['key', 'seq']))

           event
key seq         
A   0    created
    1    updated
    2    updated
    3    updated
B   0    created
    1    updated
    2    updated
C   0    created
    1    updated
    2    updated
    3    updated
    4    updated
    5    updated