Use Case: Different Attribution Models

[martinthomson]

What is the use case?

A lot of discussion at the recent meeting about attribution models.

The choice of attribution model is situational and is based on an understanding of how well the model fits the data. Attribution models measure different types of correlation. The goal is to understand - or at least estimate - causation. Lift or other techniques might be used to determine which model best fits with an advertisers goals.

A lot of attention has been focused on last touch attribution and its advantages and biases. There are lots of other models that have been contemplated, extending to a data-driven attribution that uses ML and the history of all interactions from all users to perform allocations of credit across different ad placements.

Why is it important to preserve this use case?

Advertisers being able to choose attribution model gives them a metric that better fits with their expectations and goals.

Ad Networks and Publishers being able to offer different attribution models are better able to deliver advertising that is contributes to advertiser goals.

Users probably gain nothing from this, though perhaps they are more efficiently and effectively manipulated as a result.

How is it functionally achieved today?

Existing functionality involves having access to as complete a history of interactions as possible. This enables a range of attribution models to be enabled, but it depends on computation over a lot of information that we might prefer to keep private.

Proposed Support

(I'll let other speak for the attitude/position of different proposals.)

[csharrison]

I want to outline another dimension which we discussed today, whether the platform should explicitly offer cross channel attribution. This would be a new capability that is only possible today with parties that are present on every single ad interaction.

As mentioned, there can be security issues (DOS) with supporting this, but it should be in scope when we discuss what attribution models we would like to support.

In terms of existing support, the Attribution Reporting API proposal uses a priority based system, where the highest priority impression gets full credit at conversion time. This system supports (with some caveats), first touch, last touch, and probabilistic-linear attribution. We also do not explicitly support platform-assisted cross-channel attribution.

See also #14

[alextcone]

A recent article on the subject:

https://www.adweek.com/programmatic/snap-continues-to-push-for-multi-touch-attribution-over-last-touch-last-click/

Snap’s writeup:

https://forbusiness.snapchat.com/blog/evolving-measurement-neustar