In the study's Silo-3 test, FedV-KGQA's combined candidate set contained a correct answer for 99% of MetaQA questions, 100% of PathQuestion questions and 78% of WebQSP questions. Any single silo alone reached at most 54%, 35% and 46%, respectively.
Those figures measure candidate coverage, not unrestricted answer accuracy. If the correct answer was outside the candidate set, the system could not rank it, so candidate recall bounded rankable performance.
The test used synthetic partitions
The study reports computational tests on MetaQA, PathQuestion and WebQuestionsSP, three knowledge-graph question-answering benchmarks, across 12 model configurations. Each dataset was partitioned into 3, 5 or 7 semantic silos, or separate partitions. The authors constructed those partitions; no public benchmark supplied naturally vertical multi-hop data.
FedV-KGQA concatenates the silo embeddings at a server. A frozen language encoder and a trainable two-layer multilayer perceptron project the question, which is anchored to the topic entity; candidate answers are ranked by cosine similarity.
The exchange establishes only a structural boundary: Hk leaves each silo, while server-side inference is not bounded and the paper makes no formal differential-privacy claim.
Scores varied by model and split
MRR, or mean reciprocal rank, is a measure of how highly the correct answer ranked. Under a common BERT+TransE configuration and three-silo partition, FedV-KGQA had the highest listed MRR among the adapted methods: 0.76 on MetaQA, 0.65 on PathQuestion and 0.54 on WebQSP, compared with best adapted-baseline figures of 0.71, 0.58 and 0.46.
TransE was described as the most stable of the tested knowledge-graph embedding models. On WebQSP with BERT and Silo-3, TransE's MRR was 0.54 versus 0.41 for DistMult; on MetaQA's Silo-3 test, the four KGE models' MRRs ranged from 0.71 to 0.76.
BERT and DistilBERT had similar reported performance despite BERT's 110 million parameters versus 66 million for DistilBERT. In the reported examples, RoBERTa's MRR was 0.58 versus 0.74 for BERT and DistilBERT on MetaQA's seven-silo setting, and 0.43 versus 0.54 for BERT on WebQSP's three-silo setting.
The reported MRR figures were not monotonic across silo counts. In the cited three- and five-silo BERT+TransE comparisons, MetaQA's MRR was 0.76 in one setting and 0.74 in the other; WebQSP's was 0.54 and 0.51, while PathQuestion showed a mixed pattern.
Further checks exposed limits
The three-hop test used one additional bounded expansion hop without an architectural modification. For BERT+TransE, MetaQA's MRR was 0.76 at two hops and 0.74 at three; PathQuestion's was 0.65 and 0.57. With DistilBERT+DistMult on PathQuestion, the corresponding figures were 0.66 and 0.51.
In component comparisons, FedV-KGQA's MRR was within 0.04 of the centralized upper bound on MetaQA and within 0.03 on PathQuestion and WebQSP. The no-anchoring variant recorded a PathQuestion MRR of 0.57, compared with 0.64 for the anchored version; its WebQSP H@10, the top-10 hit measure, was 0.09 lower. The no-enrichment comparison showed a 0.19 difference in WebQSP MRR and a 0.28 difference in H@10, while local-only variants recorded MRRs of 0.40, 0.20 and 0.32 on MetaQA, PathQuestion and WebQSP.
In comparisons with Gaussian embedding noise at a level of 0.05, the reported MRR values were 0.83 versus 0.79 on MetaQA, 0.68 versus 0.59 on PathQuestion, and 0.52 versus 0.48 on WebQSP. At 0.15, the reported H@10 figures were 0.87 on MetaQA and 0.86 on PathQuestion.
A benchmark result with clear boundaries
Communication estimates scaled with entity-vocabulary size rather than local triple count. To reach H@3, the top-three hit rate, of 0.35 on WebQSP, the estimate was 96.8 GB with DistilBERT and 121.1 GB with BERT. On PathQuestion, a target of 0.70 was estimated at 23.1 GB with BERT and 31.4 GB with DistilBERT; RoBERTa missed all the top targets.
The evidence is limited to computational experiments on synthetic vertical partitions of three benchmarks, not naturally vertical data or a real organizational deployment. The graph was assumed to be static, and the evaluation covered only the benchmarks' two-hop and three-hop settings.
The authors conclude that effective multi-hop question answering is achievable across split knowledge graphs, while acknowledging the static-graph setting, the absence of formal differential privacy and the fact that embedding recoverability was not measured. They list incremental updates, leakage quantification and differential-privacy mechanisms as future work.
The reported comparisons are point estimates, and no uncertainty estimates were provided.
The document is an arXiv preprint, version 1, dated 25 August 2026. It lists source code and an interactive demo under Apache License 2.0. The supplemental statement discloses Claude assistance with manuscript writing and code development, while stating that no AI tool generated the findings, tables or citations.
Paper data and sources
Original title: FedV-KGQA: Multi-Hop Question Answering over Vertically Partitioned Knowledge Graphs
Authors: Md Saikat Islam Khan Bappy, Oshani Seneviratne
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text