dmdabbs
commented
11 months ago A good topic to queue for discussion in the Wednesday weekly WICG Protected Audience call.
Open radhika-mani opened 11 months ago
dmdabbs
commented
11 months ago A good topic to queue for discussion in the Wednesday weekly WICG Protected Audience call.
Background In a previous post (https://github.com/WICG/turtledove/issues/435) Google Ads asked for having some signals about the user in reportWin so that they can be used in ML optimizations to deliver more utility from PA API. Chrome agreed with the request mentioning (https://github.com/WICG/turtledove/issues/435#issuecomment-1453921534) “we have decided to provide 12 bits of information, which equates to 4096 buckets. AdTech can choose to allocate those buckets across any number of variables they choose to.”
Deciding which variables to allocate the 12 bits to We have started using these 12 bits and have seen early signs of them helping utility. A practical challenge has been to decide how to allocate the 12 bits to different variables. A typical idea could be to try a combination of variables one at a time: Log them in reportWin, Train ML models based on them, experiment with them and iterate. This process can take a long time.
Proposal Given the tight timeline between now and 3PCD, we propose that Chrome provides extra bits prior to 3PCD to enable AdTech exploring the best use case of the 12 bits. This way AdTech can in parallel explore a few ideas and converge to delivering best utility from PA API by the 3PCD timeline. Our proposal is to have 3 times more bits (i.e. 48 bits in total, for the time before 3CPD) for exploration.