Improve datetime feature handling

Related: #1433

In AutoGluon when datetime features are detected, they are simply converted to an int.

Instead, we could generate a variety of additional features, such as:

• Year, Month, Day (#1446)
• Day of Week (#1446)
• Is Holiday / Near Holiday

Additionally, we could improve the detection of date features:

• Detect datetime in datetime_as_int form (such as 1533140820000000000, equivalent to 8/1/2018 16:27)
• Detect datetime in yyyymmdd int/str form (such as 20211222)

If you are interested in contributing this functionality, please respond to this issue.

Current datetime feature generator: https://github.com/awslabs/autogluon/blob/master/features/src/autogluon/features/generators/datetime.py

Current datetime unit test: https://github.com/awslabs/autogluon/blob/master/features/tests/features/generators/test_datetime.py

[willsmithorg]

I have a pull request + new tests ready to go. It adds ‘year’, ‘month’, ‘day’, ‘hour’, ‘minute’, ‘second’, ‘dayofweek’, ‘weekday’, ‘dayofyear’, ‘quarter’, ‘is_month_end’, ‘is_leap_year’ I think all the pd.Series.dt.* can be derived from these. I didn’t use FastAI to avoid an extra dependency.

https://github.com/awslabs/autogluon/pull/1444

[Innixma]

Thanks to @willsmithorg for contributing #1446 which adds improved datetime handling logic to AutoGluon!

@trujillo9616 if you are still interested in working on this logic, please refer to #1446 for the current state of the code, as the main logic is now implemented. An additional datetime feature that isn’t yet implemented is is_holiday.