Enhancing Forecasting in Crisis with Multimodal Federated Learning for Socio-Economic System

Venue: UAI

Reason for rejection: Reviewers found the MFL pipeline unclear and questioned the method’s novelty, lack of technical analysis, and justification for choosing SARIMAX over other forecasting methods.

Abstract

Federated learning has emerged as a powerful approach to maintaining data privacy while enriching models with diverse data sources. This paper explores the concept of multimodal federated learning (MFL), where each client possesses unique data types. We develop a pipeline for MFL and apply it to forecasting in crisis situations, aiming to improve forecast quality while maintaining overall performance. Our method employs the SARIMAX model for local training and aggregation. Experimental results demonstrate that MFL can enhance forecasting accuracy, especially during crisis periods, without sacrificing performance in normal conditions.