DOI: https://doi.org/10.36719/2789-6919/64/253-271
Attention-Guided Sparse Aggregation and Hetero Geneity-Aware Optimization for Federated Wireless Traffic Prediction
Rajakani Velumani1 
, N.Babu2 , K.Madhan3* , and P.Maithili4
Abstract. In cellular networks, wireless traffic forecasting is essential for proactive communication system adaptation. Wireless traffic prediction at the edge using Federated Learning (FL) has garnered a lot of attention due to its improved privacy protection and exemption from raw data transmission. However, existing FL-based traffic prediction methods still require substantial gradient exchange between local customers and a central server, and it remains unclear how to model the spatial dependencies among local customers within the FL architecture. To address this, we propose an FL framework, built on a Distributed Autonomous Network (DAN) architecture for 6G wireless communication, that proceeds through four stages, local modelling, gradient clipping and compression, path-level accumulation, and global optimization, to minimize communication load while maintaining forecasting accuracy. At each customer, an L2-norm gradient clipping step is followed by Top-K sparsification and 1-bit quantization to substantially shrink the size of the uploaded gradients. To offset the information lost through this compression, a Path Attention Aggregation Module learns customer-specific attention weights from each customer's upload trajectory, combining gradient-stability indicators with local performance measures, so that consistent, high-quality updates are weighted more heavily during aggregation; three customized aggregation strategies, all-correlated, δ-threshold, and k-relevant, further use gradient correlation to capture the spatial interdependence among customers. A Heterogeneity-Aware Adaptive Optimizer, built on Adam, additionally tunes each customer's learning rate, momentum, and weight-decay settings to its own traffic characteristics and enforces cross-round consistency, mitigating the effects of non-IID data across customers.
Experiments on two real-world wireless traffic databases show that the proposed approach outperforms modern algorithmic techniques and can increase communication efficacy by up to two orders of magnitude, without compromising forecasting precision.
Keywords: cellular networks, traffic forecasting, autonomous network, compression, quantization, heterogeneity