Re: [meetings] Review baseline requirements for Private Measurement from participants (#91)

Ghost ads+ is broadly used as a term and in practice. We know of about 30
firms that have used these methodologies including Google, Criteo, Pandora,
Xandr, Roku, Yahoo!, Nielsen Catalina Solutions, and many smaller firms. My
site lists a couple dozen mostly industry blog posts that explain the
concept: https://www.garjoh.com/research/.

When you run a holdout experiment where the control group does not see the
focal ad campaign, you have the challenge of identifying counterfactual
exposed users: i.e., users in the control group who would have been exposed
to the focal ad campaign if they had been assigned to the treatment group.
We propose three related methodologies:

   - Ghost ads: Tag ads/impressions in the control group where the user
   would have seen the focal ad.
   - Ghost bids: Tag ad/impressions in the control group where the focal
   advertiser would have bid on. This--together with logging bids in the
   treatment--identifies a superset of users that an advertiser wants to bid
   on.
   - Predicted ghost ads: Tag ads/impressions in the treatment and control
   group where you intended to show the ad. This is useful for situations
   where you don't fully control the ad allocation process e.g. due to
   publisher's advertiser exclusions or ad time outs. This identifies the
   superset of users that the advertiser intended to bid on. In principle you
   could do related approaches in this vein, but this is how we apply it in
   our paper.

Based on the above, I think you are describing "predicted ghost ads." If I
follow correctly, I humbly suggest calling it that because it has a
well-defined meaning and the general concept is familiar to people who run
ad experiments.

Thanks @michaelkleber!

On Mon, Jan 30, 2023 at 7:41 PM Erik Taubeneck ***@***.***>
wrote:

> @AramZS <https://github.com/AramZS> How about
>
> Can measure opportunities (an opportunity is when a user is eligible to be
> shown an ad but, due to a choice made by an advertiser-authorized system in
> order to measure that user as in a test or control group for testing, may
> have withdrawn or replaced the ad from being shown.)"
>
> I think it's fine to be wordy, but better to use the term "opportunity"
> and then define it in the parens. Also, for clarity, opportunity is
> typically a super set of impressions, so the ad may or may not end up being
> shown/withdrawn/replaced.
>
> —
> Reply to this email directly, view it on GitHub
> <https://github.com/patcg/meetings/issues/91#issuecomment-1409578122>, or
> unsubscribe
> <https://github.com/notifications/unsubscribe-auth/AM7MGWRZ7PSBXBPAEEADXD3WVBNVFANCNFSM6AAAAAAT7QHUBA>
> .
> You are receiving this because you were mentioned.Message ID:
> ***@***.***>
>


-- 
GitHub Notification of comment by garjoh
Please view or discuss this issue at https://github.com/patcg/meetings/issues/91#issuecomment-1410506138 using your GitHub account


-- 
Sent via github-notify-ml as configured in https://github.com/w3c/github-notify-ml-config

Received on Tuesday, 31 January 2023 14:53:18 UTC