MVFL: Multivariate Vertical Federated Learning for Time-Series Forecasting
Keywords: vertical federated learning, multivariate time series forecasting, resource-limited devices
TL;DR: A novel federated learning framework for multivariate time series forecasting with vertically distributed variates
Abstract: Extending multivariate time series forecasting to resource-limited devices is a critical demand for real applications, especially with the advancements in IoT technologies. A common scenario is where the variates are distributed vertically on different devices and each device needs to do local forecasting. This paper studies a resource-efficient solution for this scenario based on vertical federated learning (VFL). Prior VFL frameworks are designed for situations where only one party holds the labels and would struggle to meet the demand of the targeted scenario, as storage resources usage would increase dramatically with the number of devices. Going beyond VFL, we design multivariate vertical federated learning (MVFL) as a novel federated learning framework, where we separate communication features and local features in an embedded feature space. This design enables MVFL to utilize storage and communication resources more efficiently by eliminating the redundant models. MVFL outperforms VFL approaches in both efficiency and accuracy. On four real-world benchmarks, compared to VFL, when the storage resources are equally utilized, MVFL yields a 12.1\% relative improvement on loss with a 43\% relative improvement on communication resources usage. Even when both MVFL and VFL employ the same main model size, MVFL achieves a 75\% reduction in storage resources compared to VFL while maintaining the loss at the same level of VFL.
Primary Area: learning on time series and dynamical systems
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8592
Loading