TY  - EJOU
AU  - Feng, Yuan 
AU  - Li, Xiaotong 
AU  - Hu, Xinhua 
AU  - Zhang, Jianwei 
AU  - Cai, Zengyu 
AU  - Zhu, Liang 

TI  - A MAML-Based Federated Meta-Learning Framework for Wireless Network Traffic Prediction
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Wireless network traffic prediction enables operators to anticipate network trends, proactively develop network management strategies, and intelligently allocate network resources, thereby improving network service quality and enhancing users’ internet experience. Existing centralized network traffic prediction methods require transmitting large volumes of traffic data, which consumes significant network resources, incurs extra communication costs, and faces difficulties in full data sharing due to privacy concerns. Federated learning is a distributed learning method for multi-client joint training without sharing between clients, which can effectively solve such problems. However, the network traffic prediction method based on federated learning has problems of low accuracy due to data heterogeneity and high communication resource consumption due to the large number of federated communication rounds. To address these issues, this paper combines model-agnostic meta-learning (MAML) and proposes a MAML-based federated meta-learning framework for wireless network traffic prediction, introducing the MAML meta-learning algorithm into the client training of federated learning to enhance the learning ability of the local model and train a global model that can quickly adapt to new network traffic patterns, thereby reducing the number of federated communication rounds and communication resource consumption; during the application phase, the global model is locally fine-tuned on a held-out adaptation set at each client to obtain high-precision personalized traffic prediction models. The central server adopts a mutual information (MI) based weighted aggregation scheme to improve global model generalization. Experimental evidence shows that this approach achieves lower errors and higher accuracy with far fewer rounds, thus markedly reducing the total communication overhead.
KW  - Wireless network traffic prediction; federated learning; meta-learning; MAML; mutual information

DO  - 10.32604/cmc.2026.087978