Federated learning II

Papers

Sign in to ask or upvote questions. Sign in

1. Short paper

   Client selection under energy budgets in cross-device federated learning

   Marek Kowal (Aegean Institute of Technology) (presenting), Lina Weber (Lakeside Technical University)

   Abstract

   Battery-powered clients cannot join every round. We frame client selection as a knapsack problem that trades expected model improvement against remaining battery. In simulation with 500 phones the policy extends fleet participation time by 2.3 times with no loss in final accuracy.

   federated learning · client selection · energy · mobile devices

   Questions from the audience 2

   • ▲4

     What happens when a client's data distribution changes during training?

   • ▲2

     Is the code available, and under which licence?

2. Full paper

   Federated learning for early prediction of hospital readmission without sharing patient records

   Clara Fischer (Thalassa University) (presenting), Emil Strand (Kestrel Labs), Dora Papas (University of Eastvale)

   Abstract

   Hospitals rarely share patient data, which limits the size of training sets for readmission models. We train a gradient-boosted model across five simulated hospitals using federated learning with secure aggregation. The federated model reaches an AUC of 0.81 compared with 0.74 for the best single-hospital model. We report which features drive predictions at each site and how clinicians reviewed the explanations.

   federated learning · healthcare · secure aggregation · readmission

   Questions from the audience 2

   • ▲6

     How would the approach behave with far fewer clients, say five?

   • ▲5

     Is the code available, and under which licence?

3. Full paper

   Personalised federated models for solar generation forecasting at 15-minute resolution

   Eva Martin (Northbridge University) (presenting), Luca Greco (Aegean Institute of Technology)

   Abstract

   Rooftop solar output depends on local conditions that a single global model captures poorly. We combine a shared GRU backbone with per-site output layers trained through federated learning. Across 120 installations, day-ahead forecasting error falls by 36% compared with a centralised baseline that cannot use the private sites, and by 9% compared with a non-personalised federated model.

   federated learning · solar forecasting · personalisation · time series

   Questions from the audience 2

   • ▲3

     Did you measure the energy cost on the client devices?

   • ▲1

     Could the same idea work with vertical federated learning?