DateParts

Use DateParts for recurring calendar position, such as hour of day or day of week. Derive a Number when elapsed time, age, or recency matters.

created_at: "2026-05-28T14:30:45"

import relflow as rf

created_at = rf.DateParts(
    dateparts=["day_of_week", "hour_of_day"],
    pattern="%Y-%m-%dT%H:%M:%S",
)

Input and options

Accepts Arrow date, timestamp, or string values. Strings are parsed by Arrow using pattern when supplied; otherwise Arrow casts them to timestamps. Calendar extraction uses second-resolution timestamps. Normalize to the intended calendar timezone upstream; the field does not choose a local-time policy.

Option	Default	Meaning
dateparts	required	Nonempty, unique list of parts; casing and separators are normalized.
pattern	None	Optional strptime-style format for Arrow string parsing.
jitter	rf.Jitter()	Training noise on the generated sine/cosine coordinates.

Part	Period
day_of_year	366
week_of_year	53
month_of_year	12
day_of_month	31
week_of_month	6
day_of_week	7
hour_of_day	24
minute_of_hour	60
second_of_minute	60

Representation and output

Each part becomes a sine/cosine pair: the end of a cycle is near its beginning. relflow projects each pair and combines the results with a state embedding. Null, padded, and masked content is zero; state remains separate.

Jitter changes the circle coordinates, not the timestamp, and does not renormalize noisy pairs. Its magnitude is measured in coordinate units. Both settings of jitter.normalize use this same boundary.

Reconstruction learns state cross entropy and mean per-part cosine distance. The per-part mae metric is angular error in radians, not elapsed time.

DateParts has no decoded public content and does not generate timestamps. Use embed=True for a contextual embedding; see shared leaf options.

The month-inference proof tests a recurring relationship on withheld dates and years. The leap-boundary proof shows how selecting too few calendar coordinates can make the answer ambiguous.