Vertical Federated Learning as a Social Choice Problem

Abstract

How should a central server aggregate predictions from multiple agents, each holding partial or overlapping information, into a single collective decision — without exposing any individual agent's internal model or reasoning? Existing Federated Learning methods which satisfy our low-information requirements do not sufficiently address the online aspect of our setting. Online Learning supplies exactly this kind of guarantee, but its aggregation methods assume that some single agent is competitive on its own — an assumption that fails in the partial-feature setting Vertical Federated Learning requires. This thesis aims to close this gap by proposing methods for principled aggregation under partial feature coverage, framing this as a social choice problem. We propose a novel server strategy Trust-Based Perpetual Voting (TBPV) , which aggregates agent ballots through a scoring rule whose weights are learned online from the history of collective outcomes.