Use different Priors based on channel type

[fernaper]

Hello, first of all thank you very much for everything.

I have my pymc_marketing model working well. To do so, I'm using the following code:

from pymc_marketing.mmm import MMM, GeometricAdstock, LogisticSaturation

def generate_first_model_specification(X_train, media_columns: list[str]):
    epsilon = 1e-6

    spend_shares = (
        X_train.melt(
            value_vars=media_columns,
            var_name='channel',
            value_name='spend'
        )
        .groupby('channel', as_index=False)
        .agg({'spend': 'sum'})
        .sort_values(by='channel')
        .assign(
            spend_share=lambda x: (x['spend'] + epsilon) /
            (x['spend'].sum() + epsilon * len(x))
        )['spend_share']
        .to_numpy()
    )

    prior_sigma = spend_shares
    prior_sigma = np.clip(spend_shares, a_min=0.1, a_max=None)

    return {
        'intercept': Prior('Normal', mu=0.2, sigma=0.05),
        'saturation_beta': Prior(
            'HalfNormal', sigma=prior_sigma, dims='channel'),
        'saturation_lam': Prior('Gamma', alpha=3, beta=1, dims='channel'),
        'gamma_control': Prior('Laplace', mu=2, b=0.2),
        'gamma_fourier': Prior('Laplace', mu=0, b=1),
        'likelihood': Prior('Normal', sigma=Prior('HalfCauchy', beta=0.5)),
        'adstock_alpha': Prior('Beta', alpha=2, beta=2, dims='channel'),
    }


def build_model(X_train, media_columns: list[str], datetime_column, media_columns, control_columns, fast: bool=False) -> MMM:
    model_config = generate_first_model_specification(X_train, media_columns)
    return MMM(
        model_config=model_config,
        sampler_config={
            'progressbar': False,
        },
        time_varying_intercept=not fast,
        time_varying_media=False,
        date_column=datetime_column,
        adstock=GeometricAdstock(l_max=8),
        saturation=LogisticSaturation(),
        channel_columns=media_columns,
        control_columns=control_columns,
        yearly_seasonality=8,
    )

This work perfectly, however with certain complex datasets I'm getting really low accuracy and I need to improve it.

To do so I wanted to use different Adstock functions based on the type of channel. Therefore I tried this:

from pymc_marketing.mmm import (
    DelayedAdstock,
    MediaConfig,
    MediaConfigList,
    MediaTransformation,
    GeometricAdstock,
    LogisticSaturation,
)

def build_media_configs(media_columns: list[str]) -> MediaConfigList:
    possible_offline = {
        'tv', 'radio', 'print', 'outdoor', 'ooh', 'exteriores',
        'exterior', 'press', 'newspaper', 'magazine', 'direct mail',
        'revista', 'revistas', 'correo directo', 'correo', 'mail',
    }  # Just some offline channels to check

    offline_channels = [
        channel
        for channel in media_columns
        if channel.lower() in possible_offline
    ]

    online_channels = [
        channel
        for channel in media_columns
        if channel not in offline_channels
    ]

    media_configs = []
    if offline_channels:
        media_configs.append(
            MediaConfig(
                name='offline',
                columns=offline_channels,
                media_transformation=MediaTransformation(
                    adstock=DelayedAdstock(
                        l_max=8, delay=2).set_dims_for_all_priors('offline'),
                    saturation=LogisticSaturation().set_dims_for_all_priors('offline'),
                    adstock_first=False,
                ),
            )
        )
    if online_channels:
        media_configs.append(
            MediaConfig(
                name='online',
                columns=online_channels,
                media_transformation=MediaTransformation(
                    adstock=GeometricAdstock(
                        l_max=8).set_dims_for_all_priors('online'),
                    saturation=LogisticSaturation().set_dims_for_all_priors('online'),
                    adstock_first=True,
                ),

            )
        )

    return media_configs

The problem is that I don't know how to include those media_configs into the MMM object.

---

Can anyone help me with this?

---

Also, if you think that I'm doing something wrong, please let me know. 😄

---

Thank you very much!

[williambdean]

At the moment, the out-of-the-box MMM class doesn't support this functionality. Open to PRs!