FedGATE: Topology‐Aware Personalized Federated Learning With Utility and Carbon‐Aware Scheduling for Mobile Traffic Prediction

Abstract

Accurate mobile traffic prediction is important for proactive resource allocation, congestion management, and service assurance in fifth‐generation and emerging sixth‐generation networks. Centralized deep‐learning approaches can provide strong forecasting performance but require raw traffic data to be transferred to a common server, creating privacy, communication, and energy concerns. Federated learning (FL) avoids centralized data collection; however, Federated Averaging (FedAvg) is sensitive to non‐independent and identically distributed (non‐IID) traffic patterns and does not explicitly exploit relationships among clients. This paper presents FedGATE, a cross‐silo personalized FL framework that integrates server‐side topology‐aware aggregation, private client‐specific prediction heads, and utility and carbon‐aware participation scheduling. The server aggregates only shared‐backbone updates using graph‐constrained attention over related clients, while each client retains its prediction head locally. The scheduler adjusts participation frequency according to estimated update utility, resource cost, staleness, and carbon intensity. FedGATE is evaluated on the Telecom Italia Milan cellular traffic dataset, Edge‐IIoTset, and CIC‐IoT2023 over ten random seeds. Relative to FedAvg, FedGATE reduces mean absolute error (MAE) by approximately 8% on Milan and 20% on Edge‐IIoTset, while remaining statistically comparable to centralized training on Milan. On the Milan dataset, it reduces communication volume by approximately 37% and estimated training energy by approximately 23%. Carbon‐aware scheduling further lowers estimated carbon dioxide emissions by approximately 7% relative to the non‐carbon‐aware FedGATE variant, without a statistically significant increase in MAE. These results indicate that topology‐aware aggregation, local personalization, and resource‐aware scheduling can be integrated to improve forecasting accuracy and training efficiency in heterogeneous mobile‐network environments.