Training ML models

[lalbrizzo]

Hi @michaelkleber,

I am trying to figure out how, in this framework, one can use ML models inside the JS function. In particular, I was thinking of all the ML models that use third party cookies of users that already converted to predict which new users will be likely to convert according to their browsing navigation and I can't see how these kinds of models could be trained/used in this framework.
Do you have any idea on how these could be achieved inside this framework?
Thanks in advance,
Best
Luca

[michaelkleber]

It seems like you're asking about the possibility of Federated Learning coming to browsers! While I would love to see that in the future, the current TURTLEDOVE proposal is only about letting you run some fixed on-device JS to produce your bid; all ML training seems likely to remain on the server for now.

The Privacy Sandbox's Conversion Measurement Event-Level API is about offering servers a way to collect enough data to train conversion ML models where the objective function fits in only a few bits of data, and the Conversion Measurement with Aggregation or Aggregated Reporting APIs offer other ways to get at aggregate-level information that could be used to modify ML output.

One way this could all fit together is for the serious ML models to remain on the server and produce as output the "bidding signals" objects returned with the contextual+1p and interest-group responses, and for the in-browser JS to be a simple model combining those signals, optimized using some of those aggregation APIs.

[lalbrizzo]

Hi and thanks for the information.
So, if I have understood correctly, the main idea is to train models on servers based on the information given by the Aggregated Reporting (and/or contextual ones) and then target users using the "bidding signals" object which will be a function of the information contained in the Aggregated Reporting. Am I right?
Thanks in advance

[michaelkleber]

Yes, I think that's right — the information that goes into aggregate reporting should all be stuff that the on-device JS can use in making its bid.

[JensenPaul]

Closing this issue as it represents past design discussion that predates more recent proposals. I believe some of this feedback was incorporated into the Private Aggregation proposal. If you feel further discussion is needed, please feel free to reopen this issue or file a new issue.