Support additional DataFrame libraries

[williambdean]

Narwhals would be good route to do this. There are no dependencies so we could bring in and work

@juanitorduz mentioned this

CC: @ColtAllen

Would be good to scope what we use in comparison of what they'd support:

• DataFrame : https://narwhals-dev.github.io/narwhals/api-completeness/dataframe/
• LazyFrame : https://narwhals-dev.github.io/narwhals/api-completeness/lazyframe/

We can probably just slowly iterate to this goal.

NOTE: pandas is a dependency of arviz so our current logic would still work

Thoughts?

[williambdean]

A good one to hit would be rfm_summary

pymc-marketing/pymc_marketing/clv/utils.py

Lines 267 to 278 in 5ae96eb

	def rfm_summary(
	transactions: pandas.DataFrame,
	customer_id_col: str,
	datetime_col: str,
	monetary_value_col: str | None = None,
	datetime_format: str | None = None,
	observation_period_end: str | pandas.Period | datetime | None = None,
	time_unit: str = "D",
	time_scaler: float | None = 1,
	include_first_transaction: bool | None = False,
	sort_transactions: bool | None = True,
	) -> pandas.DataFrame:

If this can be ported to Narwhals, someone can do the aggregation in spark or duckdb

[ColtAllen]

@williambdean I definitely think this is worth implementing for clv.rfm_summary. We'll need to keep in mind that function is also called under the hood for rfm_train_test_split and rfm_segmentation.

Do you think this would also be useful in the other modeling modules?

• MMM
• Bass-Diffusion
• Customer Choice

[williambdean]

Do you think this would also be useful in the other modeling modules?

We don't provide an aggregation step in MMM or Bass-Diffusion like in CLV. Cannot speak for Customer Choice.

Allowing people to use polars instead of pandas with mmm.fit would be the goal post for me such that we can can the ModelBuilder.fit signature.

[williambdean]

I added another issue. The fivetran adaptors would be easy use for narwhals since it is just some dataframe logic

[juanitorduz]

Hey! This one is on the pipeline pymc-devs/pymc#7964 :)