Consider the following example:

from enum import Enum
import polars as pl
from typing import NamedTuple

Months = Enum(
    "MONTHS",
    [
        "Jan",
        "Feb",
        "Mar",
        "Apr",
        "May",
        "Jun",
        "Jul",
        "Aug",
        "Sep",
        "Oct",
        "Nov",
        "Dec",
    ],
)

plMonths = pl.Enum([m.name for m in Months])

plYearMonth = pl.Struct({"year": int, "month": plMonths})


class YearMonth(NamedTuple):
    year: int
    month: str


dates = pl.Series(
    "year_months",
    [YearMonth(2024, Months.Jan.name), YearMonth(2024, Months.Feb.name)],
    dtype=plYearMonth,
)

values = pl.Series("value", [0, 1])

hellos = pl.Series("hello", ["a"]*2 + ["b"]*2)

df = pl.DataFrame({
  "hello": pl.Series(["a", "a", "b", "b"]),
  "date": pl.concat([dates, dates]),
  "value": pl.concat([values, values.reverse()])
})
df

shape: (4, 3)
┌───────┬──────────────┬───────┐
│ hello ┆ date         ┆ value │
│ ---   ┆ ---          ┆ ---   │
│ str   ┆ struct[2]    ┆ i64   │
╞═══════╪══════════════╪═══════╡
│ a     ┆ {2024,"Jan"} ┆ 0     │
│ a     ┆ {2024,"Feb"} ┆ 1     │
│ b     ┆ {2024,"Jan"} ┆ 1     │
│ b     ┆ {2024,"Feb"} ┆ 0     │
└───────┴──────────────┴───────┘

I then pivot df:

pivoted = df.pivot(index="hello", columns="date", values="value")
pivoted

shape: (2, 3)
┌───────┬──────────────┬──────────────┐
│ hello ┆ {2024,"Jan"} ┆ {2024,"Feb"} │
│ ---   ┆ ---          ┆ ---          │
│ str   ┆ i64          ┆ i64          │
╞═══════╪══════════════╪══════════════╡
│ a     ┆ 0            ┆ 1            │
│ b     ┆ 1            ┆ 0            │
└───────┴──────────────┴──────────────┘

The headers have now become strings, quite understandably, as can be seen by the following unpivot:

unpivoted = pivoted.melt(id_vars="hello", variable_name="date", value_name="value")
unpivoted

shape: (4, 3)
┌───────┬──────────────┬───────┐
│ hello ┆ date         ┆ value │
│ ---   ┆ ---          ┆ ---   │
│ str   ┆ str          ┆ i64   │
╞═══════╪══════════════╪═══════╡
│ a     ┆ {2024,"Jan"} ┆ 0     │
│ b     ┆ {2024,"Jan"} ┆ 1     │
│ a     ┆ {2024,"Feb"} ┆ 1     │
│ b     ┆ {2024,"Feb"} ┆ 0     │
└───────┴──────────────┴───────┘

But recall that df‘s original date column had a struct data type plYearMonth. Is there any way at all to do the unpivot so that unpivoted‘s date data is interpreted again as plYearMonth without performing a reparsing operation?

My best guess is that this is not possible, and its instead better to have a dictionary which map between string and struct representation?