MLflow integration for BassModel

[williambdean]

Track MLflow integration for the BassModel.

What already works (free lunch)

pymc_marketing.mlflow.autolog() patches pm.sample, which BassModel.fit() calls internally. This gives us:

• Version logging (PyMC-Marketing, PyMC, ArviZ versions)
• Model diagnostics (divergences, sampling time, draws, chains)
• ArviZ summary table
• Model graph PDF
• Metadata of inputs
• InferenceData artifact

No Bass-specific code needed for any of the above.

What would need explicit support

This should build off plotting methods on BassModel and an updated notebook. The workflow would be:

import mlflow
import pymc_marketing.mlflow
pymc_marketing.mlflow.autolog(log_bass=True)

model = BassModel()

with mlflow.start_run():
    idata = model.fit(data=y)
    pp = model.sample_posterior_predictive(X=future)

    fig = model.plot_adoption_curve()
    mlflow.log_figure(fig, "adoption_curve.png")

    fig = model.plot_decomposition()
    mlflow.log_figure(fig, "decomposition.png")

Specific additions needed in pymc_marketing/mlflow.py:

1. Patch BassModel.fit — logs model config (priors for m, p, q, likelihood), model type, and saves idata.nc (pattern already used by MMM and CLV)
2. log_bass_configuration helper — logs the prior configuration for reproducibility

Dependencies

• BassModel plotting methods (for mlflow.log_figure)
• Updated notebook showcasing the full MLflow workflow

[williambdean]

Connected issues:

• Plotting methods for BassModel #2589 — Plotting methods (MLflow logs the figures)
• Update Bass notebook to use BassModel class #2590 — Updated notebook (should showcase MLflow workflow)