FedMPO: New federated graph learning handles missing modalities with 5.65% gains

A new paper from Sirui Zhang and colleagues tackles the intersection of federated learning and multimodal graph learning—a challenging scenario where data is spread across parties, and modalities (e.g., text + images on nodes) are often incomplete. Their proposed framework, FedMPO, specifically addresses two failure modes in existing two-stage pipelines (client-side completion + server-side aggregation): first, when local devices generate missing modalities without leveraging global graph semantics (topology isolation), and second, when unreliable client updates with varying modality availability skew aggregation (reliability imbalance).

FedMPO introduces three key innovations: topology-aware cross-modal generation that uses the graph’s structure to recover missing features, missing-aware expert routing that locally filters out noisy recovered signals, and reliability-aware aggregation that down-weights unreliable updates from clients. Extensive experiments across 6 benchmark datasets on 3 tasks (node classification, link prediction, etc.) show FedMPO consistently outperforms centralized and federated baselines, with gains of up to 4.10% in high-missing scenarios and 5.65% in non-IID (non-identically distributed) settings. The work is relevant for real-world applications like healthcare, social networks, and recommendation systems where data privacy and incomplete modalities are common.