PaperlayerView exact revision on arXiv

Source-linked AI summary

FedV-KGQA: Multi-Hop Question Answering over Vertically Partitioned Knowledge Graphs

Md Saikat Islam Khan Bappy, Oshani Seneviratne

arXiv:2608.24846v1cs.AI

TL;DR

FedV-KGQA targets multi-hop question answering when required facts are split across vertically partitioned organizational silos rather than available in one centralized graph. It combines local graph enrichment, silo-local KGE training, server-side embedding fusion, and topic entity anchoring while keeping raw triples and relation parameters local. Across 12 configurations and three benchmarks, it performs strongly, generalizes to 3-hop reasoning, stays within 0.03 to 0.04 MRR of the centralized upper bound, and remains robust to embedding perturbations.

  • Problem

    Existing KGQA assumes centralized graph access, while prior federated methods generally address horizontal link prediction or simpler QA rather than multi-hop reasoning over disjoint relation partitions.

  • Method

    FedV-KGQA combines local graph enrichment and KGE training with server-side entity-embedding fusion, question projection, topic entity anchoring, and silo-specific gradient return.

  • Results

    Across three benchmarks and 12 model configurations, FedV-KGQA performs strongly, generalizes to 3-hop reasoning, remains within 0.03 to 0.04 MRR of the centralized upper bound, and is robust to embedding perturbations.

  • Takeaways & Limitations

    Effective multi-hop question answering is achievable even when a knowledge graph is vertically partitioned across organizations.

  • Takeaways & Limitations

    The method establishes structural data separation but provides no formal privacy guarantee against inferences from embeddings received by an honest-but-curious server.

Abstract

from arXiv · show

Real-world data for knowledge graph question answering is often distributed across different organizations due to governance and data sovereignty constraints. While centralized systems exist, they cannot answer multi-hop questions when the required facts are split across vertically partitioned silos. In this paper, we propose FedV-KGQA, a framework for multi-hop reasoning over knowledge graphs in which organizations share entities but own disjoint sets of relations. Our approach combines local graph enrichment and knowledge graph embeddings to ensure raw triples and relation parameters never leave each silo, establishing a structural data boundary without requiring centralized graph access. We further introduce a topic entity anchoring mechanism that grounds questions in the correct graph neighborhood without any runtime inter-silo communication. We evaluate 12 model configurations across three benchmarks and show that FedV-KGQA performs strongly, remains close to centralized performance, generalizes to 3-hop reasoning, and is robust to embedding perturbations.

1 Introduction

FedV-KGQA addresses multi-hop question answering when organizations share entities but partition relation types across silos. It combines local graph and embedding processing with server-side fusion and topic anchoring while keeping raw triples and relation parameters local.

  • Motivation: Multi-hop KGQA is difficult when the facts needed for one reasoning chain are split across organizations with governance and data-sovereignty constraints.A film studio, streaming platform, and metadata aggregator may each hold different relations about the same entities.
  • Research gap: Vertical federation shares an entity space while assigning disjoint relation types to different silos, unlike horizontal federated KGE.The paper identifies this as an underexplored setting for multi-hop natural-language question answering.
  • Framework: FedV-KGQA trains local KGE models, concatenates entity embeddings at the server, projects questions into the joint space, and ranks candidates by cosine similarity.During QA training, silos receive only their gradient slices; raw triples and relation parameters remain local.
  • Framework: Topic entity anchoring grounds each question in the correct graph neighborhood without runtime inter-silo communication.The anchoring mechanism is introduced as part of the framework’s distributed reasoning design.
  • Evaluation: FedV-KGQA is evaluated across multiple KGE models, language encoders, silo configurations, and three benchmarks, showing that multi-hop QA is achievable under vertical partitioning.The evaluation covers 12 model configurations.

2 Related Work

Related work spans centralized multi-hop KGQA, federated KGE, and QA over distributed graphs. FedV-KGQA is positioned as differing from these lines in both the vertical relation split and the multi-hop natural-language answer-ranking task.

  • Multi-Hop KGQA: Centralized multi-hop KGQA assumes the full graph is available during training and inference and includes embedding-based, relational-chain, and LLM-guided methods.Examples include EmbedKGQA, UniKGQA, RelChain, RoG, Think-on-Graph, and GMeLLo.
  • Federated Knowledge Graph Embedding: Federated KGE keeps triples local but generally uses horizontal partitioning, shared relation vocabularies, and link-prediction objectives.Prior methods aggregate embeddings or relations and may add privacy, heterogeneity, unlearning, alignment, or contrastive techniques.
  • Question Answering over Distributed Knowledge Graphs: Distributed-KG QA studies include multi-hop logical queries, language-specific federated QA, and natural-language federated KGQA focused on simpler questions.The paper distinguishes these efforts from its target of multi-hop reasoning across vertically partitioned relation types.
  • Positioning: FedV-KGQA changes both the federation split and the task: relations are partitioned across silos, and answers are ranked from natural-language questions rather than predicted through link prediction.Its positioning also emphasizes reasoning chains whose hops lie in different silos.

3 Problem Formulation

The problem formulation models a knowledge graph whose entities are shared but whose relation types and triples are privately partitioned across independent silos. The task is to rank answer entities for topic-anchored questions while preserving silo-local graph and relational information.

  • Graph definition: A knowledge graph is represented as G = (E, R, T), where E contains entities, R relation types, and T relational triples.Each triple has the form (h, r, t).
  • Vertical partition: Each silo owns a private relation subset, with pairwise-disjoint subsets whose union covers the full relation vocabulary.The shared entity vocabulary is consistently identified across participants.
  • Privacy boundary: Local triple sets remain private, and the complete global triple set is never centralized or jointly accessible to any single party.This is the defining data-separation boundary in the formulation.
  • Learning task: Given a question anchored to topic entity e0, the model predicts an answer entity connected through a reasoning path of length L ≥1.Experiments focus on L ≥2, where cross-silo reasoning is required.
  • Learning task: The formulation requires ranking candidate answers while keeping each silo’s local triples and relation information, including Rk, private.The setting assumes aligned entity identifiers, benchmark-provided topic entities, schema-level relation axioms, and a static graph with precomputed candidates.

4 Methodology

FedV-KGQA combines silo-local graph enrichment and KGE with server-side embedding fusion, question projection, and topic anchoring for privacy-preserving multi-hop answer ranking. Its pipeline keeps triples and relation parameters local while returning silo-specific gradient slices during QA training.

  • Methodology: FedV-KGQA runs local graph enrichment, local KGE training, server-side fusion and QA training, then inference.
  • Phase 0: Local Graph Enrichment: Shared relation rules enrich each silo’s triples, while two-hop candidate construction combines one-hop neighbors across silos.This preserves reachability for chains whose hops lie in different silos; private-set union can hide silo-specific candidate contributions.
  • Phase 1: Local Knowledge Graph Embedding: Each silo independently trains a KGE model on enriched triples, retaining relation parameters locally and transmitting only entity embeddings.The evaluated KGE scoring functions include real-valued models with d-dimensional embeddings and complex-valued models with 2d-dimensional embeddings.
  • Phase 2: Server-Side Fusion and QA Training: The server concatenates per-silo entity embeddings and projects encoded questions into the resulting joint space for candidate answer ranking.Concatenation preserves silo-specific geometric views, whereas changing the silo set requires retraining the projection head.
  • Phase 2: Server-Side Fusion and QA Training: Topic anchoring adds the topic entity’s fused embedding to the projected question representation, directing ranking toward its graph neighborhood without runtime silo communication.
  • Phase 2: Server-Side Fusion and QA Training: Concatenated joint embeddings yield silo-specific gradient slices that the server returns to corresponding silos, while relation embeddings remain frozen.One upload-download exchange occurs per QA-training epoch; inference uses the fused embeddings, candidate set, anchored question vector, and cosine scores.

5 Experiments

The experiments evaluate FedV-KGQA across three benchmarks, six research questions, 12 model configurations, and multiple vertical silo partitions.

  • Research questions: Experiments address performance, hop generalization, component contributions, adapted baselines, embedding robustness, and communication cost.
  • Datasets: The evaluation covers MetaQA, PathQuestion, and WebQSP, which differ in domain, scale, and question complexity.
  • Metrics: FedV-KGQA reports MRR and Hits@K metrics, including H@3, H@5, and H@10, for ranking answer candidates.
  • Model configurations: Twelve configurations combine four KGE models with three frozen encoders, while only the two-layer MLP head is updated during QA training.
  • Silo configurations: Experiments use 3, 5, and 7 silos created by partitioning relation categories while sharing entities across silos.

5.3 RQ1: Performance on 2-Hop Reasoning

On 2-hop reasoning, FedV-KGQA performs consistently across KGE models, encoders, and silo counts, with dataset-specific differences in model and encoder behavior.

  • KGE model comparison: On MetaQA, all four KGE models achieve similar MRR values from 0.71 to 0.76 on the 3-silo partition.
  • KGE model comparison: With BERT on WebQSP Silo-3, TransE reaches an MRR of 0.54 while DistMult achieves 0.41.
  • Encoder comparison: DistilBERT matches BERT’s QA accuracy despite using 66M rather than 110M parameters, while RoBERTa underperforms both.
  • Encoder comparison: On MetaQA Silo-7 with TransE, RoBERTa records MRR 0.58 compared with 0.74 for BERT and DistilBERT.
  • Effect of silo count: Silo-count effects are non-monotonic: BERT+TransE on MetaQA declines from MRR 0.76 to 0.74, while some configurations improve with more silos.

5.4 RQ2: Extension to 3-Hop Reasoning

FedV-KGQA extends from 2-hop to 3-hop reasoning with a small MetaQA decline but a larger PathQuestion degradation under the 3-silo setup.

  • Evaluation design: The comparison uses the Silo-3 partition and selects each encoder’s best-performing KGE model from the 2-hop results.
  • MetaQA: On MetaQA, BERT+TransE MRR decreases by 0.02 from 0.76 to 0.74 when moving from 2-hop to 3-hop reasoning.H@10 remains 0.93 across both hop settings.
  • PathQuestion: On PathQuestion, BERT+TransE MRR drops from 0.65 to 0.57 between 2-hop and 3-hop reasoning.
  • PathQuestion: On PathQuestion, DistilBERT+DistMult MRR declines from 0.66 to 0.51 when extending to 3-hop questions.

5.5 RQ3: Ablation Study

The ablation study compares FedV-KGQA with centralized, component-removal, and local-only variants, showing that anchoring, enrichment, and federated fusion materially support answer ranking.

  • Centralized comparison: FedV-KGQA trails the centralized upper bound by 0.04 MRR on MetaQA and 0.03 on both PathQuestion and WebQSP.
  • Component ablations: Removing topic anchoring lowers PathQuestion MRR from 0.64 to 0.57 and reduces WebQSP H@10 by 0.09.
  • Component ablations: Disabling local graph enrichment reduces WebQSP MRR by 0.19 and H@10 by 0.28.The passage attributes this cost to losing inverse and chain-axiom bidirectional training signals.
  • Local-only baseline: The local-only baseline reaches MRR values of 0.40, 0.20, and 0.32 across the three datasets, indicating that individual silos lack sufficient relational knowledge.

5.6 RQ4: Baseline Comparison

FedV-KGQA is compared against adapted representative methods because no existing baseline directly operates in the vertical federated multi-hop KGQA setting. The reported comparison supports concatenation-based fusion and topic entity anchoring as an effective foundation for reasoning across silos.

  • Baseline setup: Four representative methods were adapted to the VFL setting because existing methods assume centralized access, horizontal federation, or relation aggregation.All methods used the same BERT+TransE configuration, KGE checkpoints, and 3-silo partition for comparison.
  • Reported evaluations: Figure 4 reports robustness to Gaussian noise for BERT+TransE across five silos, with σ = 0 denoting the no-noise baseline.
  • Reported evaluations: Figure 5 measures communication cost in GB needed to reach H@3 targets across encoders with TransE and three silos; dashes indicate unreached targets.
  • Comparison outcome: Concatenation-based fusion and topic entity anchoring provide a more effective foundation for multi-hop reasoning in the VFL setting.

5.7 RQ5: Robustness to Embedding Perturbations

FedV-KGQA remains relatively stable under low-level embedding noise but degrades more sharply as perturbations grow. Its communication cost is determined by embedding exchange and the number of training epochs, while the broader evaluation reports stable encoder choices and proximity to centralized performance.

  • Embedding perturbations: At σ ≤0.05, performance remains largely stable across MetaQA, PathQuestion, and WebQSP, with modest metric declines.For example, MetaQA MRR decreases from 0.83 to 0.79 at σ=0.05, while PathQuestion MRR decreases from 0.68 to 0.59.
  • Embedding perturbations: As σ increases beyond 0.10, all metrics decline more steeply, showing that stronger embedding noise reduces QA accuracy.
  • Embedding perturbations: At σ=0.15, MetaQA retains H@10 of 0.87 and PathQuestion retains H@10 of 0.86.
  • Communication cost: Communication cost scales with the entity vocabulary size, embedding dimension, number of silos, and training epochs.The per-epoch cost is Cepoch = K × |E| × d × 2 × 4 bytes, and total cost is Ctotal = T × Cepoch.
  • Overall findings: TransE with BERT or DistilBERT is the most stable configuration, and FedV-KGQA stays within 0.03 to 0.04 MRR of centralized performance.

6 Conclusion

The paper concludes that FedV-KGQA enables multi-hop question answering over knowledge graphs split across organizations without sharing raw triples or relation parameters. Across three benchmarks and 12 configurations, it generalizes to 3-hop reasoning and remains robust under embedding perturbations, while retaining important privacy and deployment limitations.

  • Conclusion: FedV-KGQA combines local KGE training, server-side embedding fusion, and topic entity anchoring for cross-silo reasoning without sharing raw triples or relation parameters.
  • Conclusion: Experiments on MetaQA, WebQSP, and PathQuestion show strong performance across 12 model configurations, 3-hop generalization, and robustness under embedding perturbations.
  • Limitations: The conclusion identifies a static-graph assumption and the absence of formal differential-privacy guarantees as key limitations.
  • Availability: The source code and interactive demo are available under the Apache License 2.0.
Loading 2608.24846v1…