Multivariate input for media data

[javiferfer]

Hello, I am currently working on constructing an MMM model, but the media data involved is multivariate in nature. To provide some context, the media features encompass not only costs but also include impressions and clicks, all of which have historical data.

Consequently, I'd like to inquire about the compatibility of your library with this type of media data input shape. To give an example, the media_data would not have the shape (117, 3) shown in demo, but it would be instead (117, 3, 3): 117: Number of weeks 3: Number of media 3: Number of features (costs, impressions, costs)

Thanks in advance!

[DiddyDaw]

In an MMM, you would typically only include the measure of impact (or influence if you like). In this case, I would go with Impressions although some would suggest clicks. Consumers don't see the cost so that features in the cost matrix but isn't the driver of behaviour (at least not directly). You only want to use the imps in this model, leave clicks altogether and use Cost in the cost matrix.

[sihohan]

Hi @DiddyDaw,

I was also trying to incorporate multiple features per media channel like @javiferfer, but from what you said it seems like this is not usually done.

As an alternative, does it make sense to pass the remainder of media features as extra features instead (e.g., use impressions as media data and clicks as extra features)?

I understand that extra features should include data that are not media data (like weather or time information), but would the setup I suggested work?

Thanks!