python - 将 Pandas 列转换为 datetime64，包括缺失值

Question

使用 Pandas 处理一些基于时间序列的数据，这些数据包含日期、数字、类别等。

我遇到的问题是让 Pandas 从 CSV 创建的 DataFrame 中正确处理我的日期/时间列。我的数据中有 18 个日期列，它们不是连续的，原始 CSV 中的未知值的字符串值为“未知”。有些列的所有单元格中都有一个有效的日期时间，并且正确地得到了 pandas read_csv 方法猜测的它们的数据类型。然而，在特定数据样本中，有些列的所有单元格都为“未知”，并且这些列被键入为对象。

我加载 CSV 的代码如下:

self.datecols = ['Claim Date', 'Lock Date', 'Closed Date', 'Service Date', 'Latest_Submission', 'Statement Date 1', 'Statement Date 2', 'Statement Date 3', 'Patient Payment Date 1', 'Patient Payment Date 2', 'Patient Payment Date 3', 'Primary 1 Payment Date', 'Primary 2 Payment Date', 'Primary 3 Payment Date', 'Secondary 1 Payment Date', 'Secondary 2 Payment Date', 'Tertiary Payment Date']
self.csvbear = pd.read_csv(file_path, index_col="Claim ID", parse_dates=True, na_values=['Unknown'])
self.csvbear = pd.DataFrame.convert_objects(self.csvbear, convert_dates='coerce')
print self.csvbear.dtypes
print self.csvbear['Tertiary Payment Date'].values

打印 self.csvbear.dtypes 的输出

Prac                            object
Doctor Name                     object
Practice Name                   object
Specialty                       object
Speciality Code                  int64
Claim Date              datetime64[ns]
Lock Date               datetime64[ns]
Progress Note Locked            object
Aging by Claim Date              int64
Aging by Lock Date               int64
Closed Date             datetime64[ns]
Service Date            datetime64[ns]
Week Number                      int64
Month                   datetime64[ns]
Current Insurance               object
...
Secondary 2 Deductible        float64
Secondary 2 Co Insurance      float64
Secondary 2 Member Balance    float64
Secondary 2 Paid              float64
Secondary 2 Witheld           float64
Secondary 2 Ins                object
Tertiary Payment Date          object
Tertiary Payment ID           float64
Tertiary Allowed              float64
Tertiary Deductible           float64
Tertiary Co Insurance         float64
Tertiary Member Balance       float64
Tertiary Paid                 float64
Tertiary Witheld              float64
Tertiary Ins                  float64
Length: 96, dtype: object
[nan nan nan ..., nan nan nan]
Press any key to continue . . .

如您所见，第三付款日期 col 应该是 datetime64 dtype，但它只是一个对象，它的实际内容只是 NaN(从字符串“Unknown”的 read_csv 函数中放置)。

我如何才能可靠地将所有日期列转换为 datetime64 作为 dtype 并为“未知”单元格设置 NaT？

Answer 1

如果您有一个全 nan 列，它不会被 read_csv 正确强制转换。最简单的就是这样做(如果列已经是 datetime64[ns] 将直接通过)。

In [3]: df = DataFrame(dict(A = Timestamp('20130101'), B = np.random.randn(5), C = np.nan))

In [4]: df
Out[4]: 
                    A         B   C
0 2013-01-01 00:00:00 -0.859994 NaN
1 2013-01-01 00:00:00 -2.562136 NaN
2 2013-01-01 00:00:00  0.410673 NaN
3 2013-01-01 00:00:00  0.480578 NaN
4 2013-01-01 00:00:00  0.464771 NaN

[5 rows x 3 columns]

In [5]: df.dtypes
Out[5]: 
A    datetime64[ns]
B           float64
C           float64
dtype: object

In [6]: df['A'] = pd.to_datetime(df['A'])

In [7]: df['C'] = pd.to_datetime(df['C'])

In [8]: df
Out[8]: 
                    A         B   C
0 2013-01-01 00:00:00 -0.859994 NaT
1 2013-01-01 00:00:00 -2.562136 NaT
2 2013-01-01 00:00:00  0.410673 NaT
3 2013-01-01 00:00:00  0.480578 NaT
4 2013-01-01 00:00:00  0.464771 NaT

[5 rows x 3 columns]

In [9]: df.dtypes
Out[9]: 
A    datetime64[ns]
B           float64
C    datetime64[ns]
dtype: object

convert_objects 不会强制将列转换为 datetime，除非它至少有 1 个非 nan 日期(这就是您的示例失败的原因)。 to_datetime 可以更具侵略性，因为它“知道”您真的想要转换它。