MOFL/D: A Federated Multi-objective Learning Framework with Decomposition

Multi-objective learning problems occur in all aspects of life and have been studied
for decades, including in the field of machine learning. Many such problems
also exist in distributed settings, where data cannot easily be shared. In recent
years, joint machine learning has been made possible in such settings through the
development of the Federated Learning (FL) paradigm. However, there is as of now
very little research on the general problem of extending the FL concept to multi-
objective learning, limiting such problems to non-cooperative individual learning.
We address this gap by presenting a general framework for multi-objective FL,
based on decomposition (MOFL/D). Our framework addresses the a posteriori
type of multi-objective problem, where user preferences are not known during
the optimisation process, allowing multiple participants to jointly find a set of
solutions, each optimised for some distribution of preferences. We present an
instantiation of the framework and validate it through experiments on a set of
multi-objective benchmarking problems that are extended from well-known single-
objective benchmarks.