jonasz / product_level_turtledove

8 stars 0 forks source link

NOTE(Oct 2020): this proposal has been incorporated into the WICG/turtledove and this repo is now archived.

To see the latest version, or in case of questions, issues, suggestions, please head to WICG/turtledove.



Product-level Turtledove*

*We focus on Turtledove for the sake of discussion, a "Product-level Sparrow" is also feasible.

Summary

In ecommerce retargeting, the majority of creatives consist of a collection of products:

creative mock

Turtledove is oblivious to the existence of products in creatives, and imposes minimal audience thresholds on the latter. This introduces a novel, fundamental difficulty in the area of useful product recommendation. Recommendation systems will have to shift from working with a single user to working with sets of users.

We argue that in ecommerce it is the product, not the creative (a collection of products), that would be the most natural conceptual unit behind Turtledove mechanisms.

If, in Turtledove, we apply the idea of minimal audience size at product level, we can:

In addition, this product-level focus brings greater clarity into areas that need to reason "what's inside the ad":

The proposed change is an extension of Turtledove, and supports all use cases that were supported originally.

Technical details

Impact on core metrics

How will the adoption of product-level Turtledove affect recommendation quality?

This is a question we haven't been able to answer for "creative-level" Turtledove. In product-level Turtledove, however, we can apply a fairly straightforward and informative analysis.

Methodology:

Caveats:

Results:

Turtledove adoption in e-commerce

In addition to core-metric impact estimates, another important factor in Turtledove adoption is research and engineering effort required.

The product-focused approach is compatible with many recommender systems employed in ecommerce retargeting today. Many state of the art systems follow the approach of scoring an item's usefulness to a single user. On the other hand, research on recommender systems working on clusters of users is scarce.

Adopting product-level focus will allow us to continue to use the recommender systems of today nearly as-is. There are some details, of course, but conceptually, we just need to get rid of very rare products and then just run the old algorithm. This greatly cuts the amount of effort required to adopt to Turtledove. The current recommender systems are a result of many years of development of teams of engineers and data scientists.

In vanilla, creative-level Turtledove, the recommender systems will likely have to be designed and built from scratch. The minimal audience threshold applied on the creative level is a difficult restriction. So far, no estimate is available on the impact it will have on recommendation quality and CTR, but we believe it would be very high.

Creative construction

In bringing creative construction to the device, we would like to retain two major objectives:

  1. Avoid enabling novel attack vectors on user's privacy.
  2. Retain the flexibility of creative formats.

We believe both goals can be met with the "strict" approach described below. We provide this approach as a minimal viable discussion starter. However, we believe there is value in seeking a more relaxed approach.

Strict approach:

This way:

A more relaxed approach would be valuable for:

Turtledove UI

On the UI level, showing the product interest groups may be much more understandable to the user.

Auditability

Turtledove assumes an audience-size approach to ensuring user privacy and comfort. By bringing more structure to the creative, an additional approach is possible:

This way:

Ad Quality

More structure in the ad makes the task of ensuring ad quality easier. SSPs would now be able to perform the task product-level.