Fed-GAME: Personalized Federated Learning with Graph Attention Mixture-of-Experts For Time-Series Forecasting

Federated learning (FL) on graphs shows promise for distributed time-series forecasting. Yet, existing methods rely on static topologies and struggle with client heterogeneity. We propose Fed-GAME, a framework that models personalized aggregation as message passing over a learnable dynamic implicit graph. The core is a decoupled parameter difference-based update protocol, where clients transmit parameter differences between their fine-tuned private model and a shared global model. On the server, these differences are decomposed into two streams: (1) averaged difference used to updating the global model for consensus (2) the selective difference fed into a novel Graph Attention Mixture-of-Experts (GAME) aggregator for fine-grained personalization. In this aggregator, shared experts provide scoring signals while personalized gates adaptively weight selective updates to support personalized aggregation. Experiments on two real-world electric vehicle charging datasets demonstrate that Fed-GAME outperforms state-of-the-art personalized FL baselines.

💡 Research Summary

Federated learning (FL) has become a key paradigm for training models on decentralized time‑series data, yet heterogeneity across clients—especially in applications such as electric‑vehicle (EV) charging demand forecasting—poses a serious challenge for both global consensus and personalized performance. Existing personalized FL (PFL) methods typically focus on local regularization, proximal terms, or clustering, while recent federated graph learning (FGL) approaches attempt to guide aggregation through a static client graph (e.g., geographic proximity). These solutions suffer from two fundamental drawbacks: (1) the client graph is predefined and cannot capture the true, task‑specific relationships among clients, and (2) the aggregation step mixes global and personalized signals into a single update vector, diluting useful information.

Fed‑GAME (Personalized Federated Learning with Graph Attention Mixture‑of‑Experts) addresses both issues with a three‑stage design. First, each client maintains a persistent private model MA and a copy of the shared global model MB. After local fine‑tuning, the client computes the parameter difference Δi = wA,i − wB and uploads it. This difference is split on the server into (i) an averaged component Δd used to update the global model (standard FedAvg‑style consensus) and (ii) a selective component Δu,i (typically the final MLP layers) that carries personalized information. Δu,i is encoded by a shared linear encoder into a low‑dimensional embedding ei, reducing communication and stabilizing subsequent attention calculations.

The core of Fed‑GAME is the Graph Attention Mixture‑of‑Experts (GAME) aggregator. It consists of M shared scoring experts Ek (simple linear networks) that evaluate pairwise relevance between client embeddings: sij,k = Ek(