FedV-KGQA: Multi-Hop Question Answering over Vertically Partitioned Knowledge Graphs
FedV-KGQA is a framework for multi-hop reasoning over knowledge graphs where organizations share entities but own disjoint sets of relations. It combines local graph enrichment and knowledge graph embeddings to keep raw triples and relation parameters within each silo, establishing a structural data boundary without centralized graph access. A topic entity anchoring mechanism grounds questions in the correct graph neighborhood without runtime inter-silo communication. Evaluation across 12 model configurations and three benchmarks shows performance close to centralized systems, generalization to 3-hop reasoning, and robustness to embedding perturbations.
The paper introduces FedV-KGQA, a framework for multi-hop question answering over vertically partitioned knowledge graphs. It addresses the challenge of answering questions when required facts are distributed across organizations with governance and data sovereignty constraints. The approach uses local graph enrichment and knowledge graph embeddings to maintain data boundaries, and a topic entity anchoring mechanism to avoid runtime inter-silo communication. Experiments demonstrate strong performance, near-centralized results, 3-hop generalization, and robustness to embedding perturbations.
FedV-KGQA leverages knowledge graph embeddings and local graph enrichment to enable multi-hop reasoning without sharing raw triples or relation parameters. The topic entity anchoring mechanism eliminates runtime inter-silo communication, which is a key technical contribution for privacy-preserving distributed reasoning. The framework's robustness to embedding perturbations suggests resilience to noise in learned representations.
This research addresses a practical need in industries where data cannot be centralized due to regulations or competitive concerns, such as healthcare, finance, and cross-organizational collaborations. By enabling multi-hop QA over distributed knowledge graphs, it could facilitate federated analytics and knowledge sharing without compromising data sovereignty.
FedV-KGQA could enable organizations to collaboratively answer complex questions without exposing sensitive data, reducing legal and compliance risks while unlocking insights from distributed knowledge. This may lower barriers to data partnerships and create new opportunities for federated AI services.
Potential next steps include scaling to larger knowledge graphs, extending to more complex reasoning patterns, and evaluating on real-world federated datasets. The approach may inspire further work on privacy-preserving graph neural networks and federated knowledge graph completion.