v1.0 Release Checklist

[ColtAllen]

Repo Maintainers: Please update this checklist and link to appropriate issues per internal discussions.

Contributors & Users: We appreciate hearing your thoughts!

MMM

• Deprecate legacy model class in favor of MultiDimensionalMMM

CLV

• API change: CLVModel(data,model_config).fit() -> CLVModel(model_config).fit(data)
  • Consolidate CLV load with ModelBuilder classmethod #1380
  • Option to exclude fit data when saving models #1356
  • Create a generic test for save/load for CLV classes #306
  • Better documentation of save and load methods #834
• sBG Notebook
• BUG: ShiftedBetaGeoModel predictive method parametrization #2114
• BG/BB Notebook
• Consolidate fit methods into ModelBuilder for cross-module support
• CLV Deprecations #1516 (requires expansion)
• Expand mlflow Support
  • complete: Add autologging for CLV models #1354
  • related: Add CLVWrapper #1377

Customer Choice

• Model Builder inheritance for save/load and expanded fit functionality? (requires API change)

Bass Diffusion

• Model Builder inheritance for save/load and expanded fit functionality (Use ModelBuilder for Bass Model #2110)

[williambdean]

Isn't the customer choice model already supporting save / load. I see this

pymc-marketing/pymc_marketing/customer_choice/mv_its.py

Lines 92 to 127 in 4983c1e

	def create_idata_attrs(self) -> dict[str, str]:
	"""Create the attributes for the InferenceData object.
	Returns
	-------
	dict[str, str]
	The attributes for the InferenceData object.
	"""
	attrs = super().create_idata_attrs()
	attrs["existing_sales"] = json.dumps(self.existing_sales)
	attrs["saturated_market"] = json.dumps(self.saturated_market)
	return attrs
	@classmethod
	def attrs_to_init_kwargs(cls, attrs) -> dict[str, Any]:
	"""Convert the attributes of the InferenceData object to the __init__ kwargs.
	Parameters
	----------
	attrs : dict
	The attributes of the InferenceData object.
	Returns
	-------
	dict
	The __init__ kwargs for the class.
	"""
	return {
	"model_config": json.loads(attrs["model_config"]),
	"sampler_config": json.loads(attrs["sampler_config"]),
	"existing_sales": json.loads(attrs["existing_sales"]),
	"saturated_market": json.loads(attrs["saturated_market"]),
	}

CC: @NathanielF

[ColtAllen]

Isn't the customer choice model already supporting save / load.

It isn't being tested in the Logit models, and the internal changes I have in mind for the CLV models may break that functionality.

@NathanielF how large are the datasets typically for the Logit models? The reason I'm redoing the CLV API is because those models sometimes rival open-sourced LLMs in size due to the massive datasets, creating I/O issues when logging to MLFlow. The only way to fix this is to have the option to exclude fit data when save/loading, which isn't possible when the model is initialized with data.

[NathanielF]

@williambdean on the mvits model i believe @drbenvincent originally implemented that class, but i think i substantially copied the class design for the MnLogit and Nested logit classes.

@ColtAllen on the question of size it's likely worth distinguishing between the mvits model and the nestedl logit and mnl logit. For the latter two, the data sets i'd expect are quite small ussually survey response data. The data itself is typically expensive to gather so it's hard to accumulate huge amounts of data. Think of the effort required to accurately describe your own company's product + your major competitors in a market.

For the mvits that takes timeseries data of product sales and this can be easier to gather, but you still need accurate sales data for your competitors. Not trivial. But i'd expect medium to large data sets for MVITs models. @drbenvincent might be able to say more....

[ColtAllen]

Thanks @NathanielF, that's helpful. The only advantage I see then in changing the Logit API is for library standardization, plus it would clean up the internals of both models.

it's likely worth distinguishing between the mvits model and the nestedl logit and mnl logit.

Both Logit models are initialized with fit data, just like the CLV models. The MVITS model follows the MMM convention and does not require modification.