Enforce domain checks for media transformations

[daniel-saunders-phil]

Checklist:

Adstock Transformations:

• geometric adstock Enforce check_parameters for alpha in geometric_adstock #960
• ...

Saturation Transformations

• ...

Any other parameters that would have a restricted domain?

Original message

The docstring says alpha must be between 0 and 1. But you can pass negative and larger alpha values and it will produce strange results, silently. Negative alpha subtracts impressions from each day. Adstock greater than 1 causes exponential accumulation of impressions over time. I can't see users ever finding this behaviour helpful.

import numpy as np
from pymc_marketing.mmm.transformers import geometric_adstock

geometric_adstock(np.array([100,200,50]),alpha = 1.1).eval()
geometric_adstock(np.array([100,200,50]),alpha = -0.1).eval()

Proposed solution: Adding a check would make it easier to spot mistakes, especially in large hierarchical models with complex parameterizations.

    """
    if np.any(alpha >= 1) or np.any(alpha < 0):
            raise ValueError("alpha must be between greater than or equal to 0 and less than 1.")

    w = pt.power(pt.as_tensor(alpha)[..., None], pt.arange(l_max, dtype=x.dtype))
    w = w / pt.sum(w, axis=-1, keepdims=True) if normalize else w
    return batched_convolution(x, w, axis=axis, mode=mode)

[wd60622]

Think this is a good idea. Probably want to use pymc / pytensor tooling though for the checks instead of numpy. The pymc.logprob.utils.CheckParameterValue can be used.

Example:
https://github.com/pymc-devs/pymc-experimental/blob/af91b423095eb37017eb84ded91c14f0eae08494/pymc_experimental/distributions/continuous.py#L331

[wd60622]

Thoughts @juanitorduz ?

[juanitorduz]

Yes! We should use pytensor.