自定义带_metadata的pandas Series子类，索引时如何保留序列名称？

在pandas 2.2.3版本中，我们定义两种pandas.Series的自定义子类变体：第一种未添加自定义属性，第二种通过_metadata指定自定义元数据。

代码示例 #

import pandas as pd


class MySeries(pd.Series):

    @property
    def _constructor(self):
        return MySeries


seq = MySeries([*'abc'], name='data')

print(f'''Case without _metadata:
  {isinstance(seq[0:1], MySeries) = }
  {isinstance(seq[[0, 1]], MySeries) = }
  {seq[0:1].name = }  
  {seq[[0, 1]].name = }
''')

class MySeries(pd.Series):

    _metadata = ['property']

    @property
    def _constructor(self):
        return MySeries


seq = MySeries([*'abc'], name='data')
seq.property = 'MyProperty'

print(f'''Case with defined _metadata:
  {isinstance(seq[0:1], MySeries) = }
  {isinstance(seq[[0, 1]], MySeries) = }
  {seq[0:1].name = }  
  {seq[[0, 1]].name = }
  {getattr(seq[0:1], 'property', 'NA') = }  
  {getattr(seq[[0, 1]], 'property', 'NA') = }  
''')

运行输出 #

Case without _metadata:
  isinstance(seq[0:1], MySeries) = True
  isinstance(seq[[0, 1]], MySeries) = True
  seq[0:1].name = 'data'  
  seq[[0, 1]].name = 'data'

Case with defined _metadata:
  isinstance(seq[0:1], MySeries) = True
  isinstance(seq[[0, 1]], MySeries) = True
  seq[0:1].name = 'data'  
  seq[[0, 1]].name = None         <<< Problematic result of indexing
  getattr(seq[0:1], 'property', 'NA') = 'MyProperty'  
  getattr(seq[[0, 1]], 'property', 'NA') = 'MyProperty'  

可见，定义_metadata后，切片操作可保留序列名称，但列表索引操作会丢失名称；未定义_metadata时两种操作均保留名称。

解决方案 #

方法1：将name加入_metadata列表 #

pandas在处理自定义元数据时，会自动复制_metadata中声明的属性到新实例。将name添加到_metadata中，就能确保索引操作后name属性被保留：

class MySeries(pd.Series):
    # 将name加入自定义元数据列表
    _metadata = ['property', 'name']

    @property
    def _constructor(self):
        return MySeries


seq = MySeries([*'abc'], name='data')
seq.property = 'MyProperty'

print(f'''Case with name in _metadata:
  {isinstance(seq[[0, 1]], MySeries) = }
  {seq[[0, 1]].name = }
  {getattr(seq[[0, 1]], 'property', 'NA') = }
''')

运行后输出：

Case with name in _metadata:
  isinstance(seq[[0, 1]], MySeries) = True
  seq[[0, 1]].name = 'data'
  getattr(seq[[0, 1]], 'property', 'NA') = 'MyProperty'

方法2：重写__getitem__方法手动保留名称 #

如果不想修改_metadata，可以重写__getitem__方法，在返回新实例时检查并设置name：

class MySeries(pd.Series):
    _metadata = ['property']

    @property
    def _constructor(self):
        return MySeries
    
    def __getitem__(self, key):
        result = super().__getitem__(key)
        # 仅当结果是MySeries实例且name为空时，继承原实例的name
        if isinstance(result, MySeries) and result.name is None:
            result.name = self.name
        return result


seq = MySeries([*'abc'], name='data')
seq.property = 'MyProperty'

print(f'''Case with overridden __getitem__:
  {isinstance(seq[[0, 1]], MySeries) = }
  {seq[[0, 1]].name = }
  {getattr(seq[[0, 1]], 'property', 'NA') = }
''')

运行后同样能得到保留name的结果。

内容的提问来源于stack exchange，提问作者vitalizzare