FedV-KGQA in Practice: Design Lessons and an Interactive Prototype
Md Saikat Islam Khan Bappy, Oshani Seneviratne
cs.AI · cs.CL · cs.IR
Quick verdict
ISWC poster distilling FedV-KGQA: federated fusion nears centralized accuracy on 3 KGQA benchmarks; honest scope, no significance tests or privacy proof.
Summary
This ISWC 2026 companion poster distills cross-experiment lessons from FedV-KGQA, a system for multi-hop question answering over vertically partitioned knowledge graphs in which each silo enriches its local triples and trains a knowledge graph embedding (TransE, DistMult, ComplEx, or RotatE), while a server concatenates the silo entity embeddings, projects a frozen-encoder question vector, anchors it at the benchmark-supplied topic entity, and ranks candidates by cosine similarity. On MetaQA, PathQuestion, and WebQSP across three to seven silos, federated fusion trails a centralized bound by 0.03–0.04 MRR while the best single silo collapses to 0.40/0.20/0.32 MRR, beating adapted FedE and RelChain baselines by up to 0.08 and 0.12. Ablations show anchoring and enrichment matter more than encoder choice, and the cheapest encoder depends on the target Hits@3. Single-silo candidate coverage (54/35/46%) versus fused (99/100/78%) mechanistically explains the gap. An interactive prototype traces real inference per question on released checkpoints. Limitations: the privacy boundary is structural, not formal — entity embeddings and two-hop candidate sets reach the server; no significance testing or variance is reported; silo count is confounded with relation composition; the prototype covers one configuration (TransE+BERT, three silos, two-hop MetaQA); sparse-update savings are unmeasured.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Federated fusion by concatenating silo entity embeddings recovers most of centralized multi-hop KGQA accuracy (within 0.03–0.04 MRR on MetaQA, PathQuestion, WebQSP), while the best single silo recovers little (0.40/0.20/0.32 MRR).
- Topic-entity anchoring and local rule-based enrichment contribute more to accuracy than the choice of KGE model or question encoder.
- The cheapest encoder depends on the target accuracy, not parameter count: DistilBERT is cheaper at Hits@3 0.35 on WebQSP (96.8 vs 121.1 GB) but BERT is cheaper at Hits@3 0.70 on PathQuestion (23.1 vs 31.4 GB).
- The privacy boundary is structural only: raw triples and relation embeddings stay local, but entity embeddings and two-hop candidate sets reach the server and may leak information.
Strengths
Honest scoping and self-limitation throughout
Section 2's Scope paragraph excludes aggregation/counting/temporal questions and entity linking; Section 3 states the WebQSP ceiling (0.78) from fused candidate coverage; Finding 3 calls the noise study 'a robustness analysis, not a privacy guarantee'; sparse updates are flagged 'We have not measured this'.
Mechanistic coverage analysis explains why single silos fail
Section 3 shows single-silo candidate sets cover only 54%/35%/46% of gold answers on MetaQA/PathQuestion/WebQSP versus 99%/100%/78% fused — attainable accuracy is capped before ranking starts, converting the headline finding into an explanation.
Genuine cross-experiment synthesis, not a repackaging
Artifact-backed reproducibility including a live failure-adjacent trace
Well-differentiated positioning
Weaknesses
No statistical rigor for small-margin claims
Privacy framing without any leakage quantification
Confounded design variables and deferred experimental detail
Prototype demonstrates one narrow configuration
Cost story left incomplete
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: FedV-KGQA in Practice (arXiv:2609.13661)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is FedV-KGQA?
A federated framework for answering multi-hop questions over a knowledge graph that is split vertically across organizations: each party keeps its triples and relation embeddings local, trains its own knowledge graph embedding, and only entity embeddings are shared with a server that concatenates them to rank answer entities for a natural-language question.
How close does FedV-KGQA get to centralized accuracy?
In the paper's experiments on MetaQA, PathQuestion, and WebQSP, the federated version trails a centralized upper bound by about 0.03–0.04 MRR, while the best single silo without fusion reaches only 0.40/0.20/0.32 MRR respectively on those benchmarks.
Is FedV-KGQA privacy-preserving?
Only structurally: raw triples and relation embeddings never leave a silo. But entity embeddings and precomputed two-hop candidate sets do reach the server, and the authors explicitly state this is not a formal privacy guarantee and that leakage from these artifacts is an open question.
Where can I get the code or try the prototype?
The authors release source (github.com/brains-group/fedv-kgqa-source) and an interactive prototype (github.com/brains-group/fedv-kgqa-prototype) that runs real inference on released checkpoints — specifically TransE+BERT with three silos on two-hop MetaQA — tracing every pipeline stage for a chosen question.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 2 is an image; its per-point values (panels a–d) are only available as quoted in the prose and caption, so fine-grained curve shapes could not be verified.
- Hyperparameters, baseline-adaptation details, and full per-benchmark results are deferred to the companion paper [arXiv:2608.24846], which is not included in the provided text; several judgments (baseline fairness) rest on this poster alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.