As organizations increasingly collaborate while keeping their data separate for privacy and governance reasons, answering complex questions that require combining information from multiple sources has become a major challenge. A newly published research paper introduces FedV-KGQA, a novel approach that allows different organizations to jointly answer multi-step questions over distributed knowledge graphs—structured data that maps relationships between entities—without sharing their raw data. This breakthrough could help industries like healthcare, finance, and government extract richer insights from fragmented data while respecting privacy constraints.
Key Takeaways
- FedV-KGQA enables multi-hop question answering over vertically partitioned knowledge graphs, where entities are shared but relations are split across organizations.
- The system keeps raw data and relation details within each organization’s silo, ensuring data privacy and compliance with governance rules.
- By combining local graph enrichment and knowledge graph embeddings, the method approximates centralized system performance without needing a unified data store.
- FedV-KGQA supports complex reasoning up to three hops (steps) and is robust against noise in the data representations.
Knowledge graphs represent facts as entities (like people, places, or things) connected by relations (such as “works at” or “located in”). In many real-world scenarios, different organizations hold different pieces of these graphs. For example, one hospital might have patient data linked to diagnoses, while another holds treatment records. Answering a question that spans both requires multi-hop reasoning—following a chain of relationships across these separate graphs. Traditional centralized systems struggle here because they need access to all the data, which is often impossible due to privacy or legal restrictions.
FedV-KGQA tackles this problem by operating in a federated setting where each organization retains control over its portion of the knowledge graph. The “vertical partitioning” means that while all parties know about the same entities, the relationships between them are split. The researchers’ approach uses two main techniques to enable joint reasoning without data sharing. First, local graph enrichment allows each silo to enhance its graph with additional inferred connections based on its own data. Second, knowledge graph embeddings transform entities and relations into mathematical vectors that capture their meaning and connections in a way that can be processed by machine learning models without revealing the original data.
To efficiently connect questions to the right parts of these distributed graphs, the team introduced a topic entity anchoring mechanism. This technique identifies the key entity a question is about and “grounds” the reasoning process in the relevant neighborhood of the graph within each silo, avoiding the need for back-and-forth communication during question answering. The researchers tested twelve different model setups on three benchmark datasets, showing that FedV-KGQA performs nearly as well as centralized systems that have full data access. It also generalizes to questions requiring reasoning across up to three relational hops and remains stable even when the embeddings are slightly perturbed.
This research offers a promising path forward for industries where data privacy and sovereignty are paramount but collaborative insights are still needed. By enabling multi-hop reasoning over distributed knowledge graphs without centralizing sensitive data, FedV-KGQA could facilitate better decision-making in fields like healthcare diagnostics, financial fraud detection, and cross-agency government analysis. Future work may explore scaling the approach to even larger networks of organizations and more complex reasoning tasks, as well as integrating additional privacy-preserving technologies to further strengthen data security.
Based on research published on arXiv by Md Saikat Islam Khan Bappy, Oshani Seneviratne.