Knowledge graphs are increasingly distributed across autonomous organizations that share an entity space but own disjoint subsets of relations, creating a vertical partition. Multi‑hop queries often need to combine facts from several silos, making the partitioning strategy a key decision that impacts communication cost, index size, load balance, and query latency.
We formalize vertical partitioning as a design space and compare four strategies: semantic‑domain grouping, frequency‑balanced partitioning, co‑occurrence graph‑cut partitioning, and random partitioning. Five metrics are used for evaluation: communication cost, candidate index size, cross‑silo path length, load balance, and end‑to‑end query latency.
The results show that three of these metrics are determined mainly by the underlying graph structure and the number of silos, not by the specific partitioning method. Consequently, the design problem reduces to a trade‑off between cross‑silo path length and load balance.
Experiments on MetaQA and PathQuestion employ a fixed federated knowledge‑graph QA architecture based on TransE embeddings and a frozen BERT encoder, tested under three silo configurations. By keeping the learning model unchanged, we isolate the effect of partitioning. The trade‑off between locality and balance is pronounced only when each silo can hold multiple relations, and it weakens as the number of silos grows.
The study offers practical guidance for deployments constrained by cross‑silo reasoning or silo load.
Review