Towards Modality-imbalanced Federated Graph Learning: A Data Synthesis-based Approach

Researchers have proposed two new approaches, FedMGS and FedEPD, to address challenges in federated graph learning, particularly modality imbalance and long-tailed data distributions.

Federated Graph Learning (FGL) is a collaborative training paradigm that preserves data privacy across distributed clients^[2]. However, it faces challenges such as modality imbalance and long-tailed data distributions. Modality imbalance occurs at two levels: client-level, where certain clients lack entire modalities, and node-level, where individual nodes exhibit missing attributes^[1]. Existing methods attempt to address these issues through topology-agnostic statistical compensations but often fail under data scarcity^[2]. To overcome these limitations, researchers have proposed FedMGS, a data synthesis-based approach that integrates three core components: an availability-aware graph encoder, a prototype-guided latent semantic synthesizer, and a reliability-calibrated semantic fusion mechanism. FedMGS has been shown to consistently outperform competitive baselines with gains up to 17.41%^[1]. Another approach, FedEPD, separates topological purification from semantic recalibration and has achieved state-of-the-art performance across diverse long-tailed benchmarks^[2].