FedV-KGQA is a framework designed to enable multi-hop question answering over knowledge graphs that are distributed across different organizations. In this setting, organizations share the same set of entities but own disjoint sets of relations, preventing any single party from accessing the full graph required to answer complex, multi-hop questions. FedV-KGQA allows these organizations to collaborate on answering questions without sharing their raw data or relation parameters.
Addressing Distributed Data Silos
Real-world knowledge is often fragmented. For example, one organization might hold data about film directors, while another holds data about actors. Answering a question that requires connecting these two facts—such as identifying actors who starred in films directed by a specific person—typically requires access to the entire graph. Centralizing this data often violates governance, commercial, or data sovereignty constraints. FedV-KGQA addresses this by allowing organizations to keep their raw triples and relation parameters local, establishing a structural data boundary while still enabling the system to perform multi-hop reasoning.
How the Framework Works
The FedV-KGQA process consists of four main phases:
Local Graph Enrichment: Each silo uses a shared set of relation rules to enrich its local data, such as adding inverse properties or chains of relations. This improves the quality of entity representations and ensures that answer entities are reachable within the silo's local graph.
Local Embedding Training: Silos independently train knowledge graph embedding models on their local data. Only the resulting entity embedding matrices are sent to a central server.
Server-Side Fusion and Training: The server concatenates the entity embeddings from all silos into a joint representation. It then uses a topic entity anchoring mechanism to ground the question in the correct neighborhood of the graph. This allows the system to rank candidate answers without needing further communication with the silos during runtime.
Gradient Return: During training, the server calculates gradients and returns specific slices to the corresponding silos, allowing them to update their local embeddings without exposing their private data.
Performance and Capabilities
Researchers Md Saikat Islam Khan Bappy and Oshani Seneviratne evaluated FedV-KGQA across 12 model configurations and three benchmarks. The results indicate that the framework performs strongly, achieving results close to those of a centralized system. The authors report that the approach generalizes to 3-hop reasoning and remains robust to embedding perturbations.
Considerations and Limitations
While FedV-KGQA provides a structural data boundary, the authors note that it does not provide a formal privacy guarantee. The framework assumes that silos share a consistently aligned entity identifier space and that the graph remains static. Additionally, because the server uses a concatenation-based fusion method, the model is not invariant to silo permutations; adding or removing a silo requires retraining the projection head. The system is designed for scenarios where relation types are disjoint across organizations, and it does not address cases where relations overlap.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!