Scalable Vertical Federated Learning via Data Augmentation and Amortized Inference

• DOI: 10.48550/arxiv.2405.04043
• Publication Date: 2024-05-07
• Authors: Conor Hassan; Matt Sutton; Antonietta Mira; Kerrie Mengersen
• Keywords: Amortized analysis; Federated Learning

Vertical federated learning (VFL) has emerged as a paradigm for collaborative model estimation across multiple clients, each holding distinct set of covariates. This paper introduces the first comprehensive framework fitting Bayesian models in VFL setting. We propose novel approach that leverages data augmentation techniques to transform problems into form compatible with existing algorithms. present an innovative formulation specific scenarios where joint likelihood factorizes product client-specific likelihoods. To mitigate dimensionality challenge posed by augmentation, which scales number observations and we develop factorized amortized variational approximation achieves scalability independent observations. showcase efficacy our through extensive numerical experiments on logistic regression, multilevel hierarchical split neural net model. Our work paves way privacy-preserving, decentralized inference vertically partitioned scenarios, opening up new avenues research applications various domains.