Would raw dates work in lieu of a combination of year and julian date?
As the project is currently structured, the raw data is read in with a date field with the format YYY_MM_DD.
When the training set is being prepared, this date is exchanged for two new fields, $year and $jdate which are both ...ints?
Those are then normalized as a step in the model recipe. But what if I skipped the above step, kept the $date field, and normalized that instead? Would recipes know how to handle that? And would that potentially have the advantage of increasing nearest neighbor estimates at boundaries -- helping the algorithm to connect, say December of one year with January of the next year?
Would raw dates work in lieu of a combination of year and julian date?
As the project is currently structured, the raw data is read in with a date field with the format YYY_MM_DD.
When the training set is being prepared, this date is exchanged for two new fields,
$yearand$jdatewhich are both ...ints?Those are then normalized as a step in the model recipe. But what if I skipped the above step, kept the $date field, and normalized that instead? Would recipes know how to handle that? And would that potentially have the advantage of increasing nearest neighbor estimates at boundaries -- helping the algorithm to connect, say December of one year with January of the next year?