CD2Linear: Convolution-based Federated Learning for Efficient Time Series Forecasting in the Metaverse

Federated Learning (FL) is a powerful framework for collaborative machine learning, particularly beneficial in the Metaverse context where preserving data privacy and optimizing performance on resource-constrained edge devices are crucial. In immersive Metaverse applications, accurate time series prediction is vital for maintaining seamless user experiences, real time responsiveness, and personalized services, as well as for efficient resource scheduling in highly dynamic virtual environments. In this paper, we introduce CD2Linear, a novel convolution-based decomposition method tailored explicitly for time series forecasting within federated learning scenarios relevant to Metaverse applications. By employing learnable depth-wise convolutional filters, CD2Linear significantly enhances predictive accuracy and adaptability over traditional decomposition approaches. Extensive experimental evaluations demonstrate that our method achieves superior forecasting performance while ensuring computational efficiency, effectively addressing critical challenges such as scalability, privacy, latency, and limited computational resources inherent to Metaverse computing and networking.