Federated Reconstruction for Matrix Factorization in OpenFL

[mansishr]

Federated reconstruction utilizes partially local federated learning where some client parameters are never aggregated on the server. Tensorflow Federated has an example and a blogpost on federated reconstruction using MovieLens 1M data. Users can review the example and add support for federated reconstruction in OpenFL using Workflow API.

[bjklemme-intel]

I would like to work on this issue.

[bjklemme-intel]

@mansishr I reviewed the paper, blogpost, and tensorflow example. The implementations mentioned in the blogpost (and also the google open source library) all use Tensorflow Federated, which is another federated learning platform. So I can't import the functions from there.

It looks like this could be implemented as an example in the Workflow API but I would have to write implementations of the reconstruction algorithm as well as reproduce the workflow for federated reconstruction in the collaborator nodes. For maximum code re-use, would it not be better to encapsulate the functionality into an API instead of embedding it into an example?

Also, please clarify what you want the issue to accomplish. The "user" is mentioned but I don't think a user would be able to do this without the example or API already written. (That is what I am proposing to do.) Did you mean the developer working on the issue?