FedCAG: Federated Causality-Aware Graph Learning for Multi-Cloud Workload Forecasting

Recent large-scale outages at major cloud providers such as AWS and GCP have exposed the fragility of relying on a single cloud. To improve resilience, fault tolerance, and business continuity, many enterprises are moving their services to multi-cloud environments. However, multi-cloud deployment also makes workload forecasting substantially harder. Existing multi-cloud workload forecasting methods are typically designed for either centralized or isolated environments, leading to risks of private data leakage or limited generalization. To address these challenges, we propose \textbf{FedCAG}, a \textbf{Fed}erated \textbf{C}ausality-\textbf{A}ware \textbf{G}raph learning paradigm for multi-cloud workload forecasting. Instead of treating federated learning as parameter averaging over local predictors, FedCAG jointly federates predictive models and graph-structured dependency priors. Each client constructs causality-aware, spatial, and temporal graphs from its private telemetry to model directed inter-metric influence, metric-level interactions, and intra-window temporal dynamics. To address strong cross-cloud heterogeneity, FedCAG further combines causality-aware representation fusion, adaptive graph refinement, and client-specific personalization, enabling the global model to benefit from shared workload structures while preserving local specificity. Experiments on real-world datasets show that FedCAG consistently outperforms mainstream federated baselines and even several centralized methods trained on the full dataset, delivering stable and accurate forecasting across heterogeneous and privacy-constrained deployments. The source code is available at: \url{https://anonymous.4open.science/r/FedCAG-FB9C}.