NEWS · RESEARCH · #131
Cost Characterization of Vertically Partitioned Federated Knowledge Graphs (arXiv)
This arXiv cs.AI paper formalizes vertical partitioning of federated knowledge graphs and compares four partitioning strategies (semantic domain grouping, frequency-balanced, co-occurrence graph-cut, and random) across five metrics: communication cost, candidate index size, cross-silo path length, load balance, and query latency. Using a fixed federated QA architecture (TransE embeddings and a frozen BERT encoder) on MetaQA and PathQuestion benchmarks and three silo configurations, the authors find that three of the five metrics are determined by the graph structure and silo count, reducing the design trade-off primarily to cross-silo path length versus load balance, and that the locality-versus-balance trade-off weakens as silo count increases.
KEY POINTS
- This arXiv cs.AI paper formalizes vertical partitioning of federated knowledge graphs and compares four partitioning strategies (semantic domain grouping, frequency-balanced, co-occurrence graph-cut, and random) across five metrics: communication cost, candidate index size, cross-silo path length, load balance, and query latency.
- Using a fixed federated QA architecture (TransE embeddings and a frozen BERT encoder) on MetaQA and PathQuestion benchmarks and three silo configurations, the authors find that three of the five metrics are determined by the graph structure and silo count, reducing the design trade-off primarily to cross-silo path length versus load balance, and that the locality-versus-balance trade-off weakens as silo count increases.
- The paper clarifies which cost factors depend on partitioning versus inherent graph/silo parameters and identifies the central trade-off (cross-silo path length vs. load balance), providing actionable guidance for designing federated KG deployments that require cross-silo reasoning or balanced load.
WHY IT MATTERS
The paper clarifies which cost factors depend on partitioning versus inherent graph/silo parameters and identifies the central trade-off (cross-silo path length vs. load balance), providing actionable guidance for designing federated KG deployments that require cross-silo reasoning or balanced load.