FedV-KGQA in Practice: Design Lessons and an Interactive Prototype

Details

Multi-hop knowledge graph question answering typically assumes a single system can access the entire graph. In practice, organizations share entity identifiers but own disjoint relation subsets — a vertical partition where no single party sees a complete reasoning chain.\n\nFedV-KGQA addresses this by having each silo train a local knowledge graph embedding on its own triples. A central server concatenates silo-specific entity views, anchors the projected question at the topic entity, and ranks candidates by similarity. Raw triples and relation embeddings never leave individual silos.\n\nThe key empirical findings: federated fusion recovers most of the centralized accuracy while a single silo recovers little — confirming that the collaborative approach is necessary. Anchoring and enrichment strategies matter more than the choice of embedding model. The cheapest encoder that meets target accuracy depends on the accuracy requirement, not parameter count.\n\nThis is a poster paper with a live interactive prototype, so results are preliminary and limited in scope. It offers a useful architecture template for organizations that need to reason over shared entity spaces without data sharing, such as healthcare networks or financial consortia.