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

Key Findings

Methodology

FedV-KGQA combines local graph enrichment with relation rule expansion and entity anchoring. Each silo trains local knowledge graph embeddings (e.g., TransE, DistMult) without sharing triples. Shared relation rules, such as inverse and chain axioms, enhance local graph connectivity. The server fuses entity embeddings via concatenation, projects questions with pre-trained transformers (BERT, RoBERTa), and anchors questions to topic entities. During training, only entity embeddings are exchanged, preserving data privacy. The approach supports multi-hop reasoning by ensuring cross-silo paths through rule-based graph expansion and anchoring, enabling answer ranking without raw data exchange.

Key Results

On MetaQA, WebQSP, and PathQuestion datasets, models achieved up to 0.76 MRR and 0.88 Hits@3, outperforming baseline federated methods. The models effectively supported 3-hop reasoning, with high coverage of gold answers within candidate sets. Experiments demonstrated robustness to embedding perturbations and low communication overhead, validating practical deployment potential.
Disjoint relation subset partitioning across silos maintained high answer coverage (up to 99%) and accuracy, confirming the effectiveness of rule-based graph expansion and anchoring. The models maintained performance across different embedding algorithms and question encoders, indicating generalizability.
Ablation studies showed relation rule expansion and topic anchoring significantly improved multi-hop reasoning, especially in complex, multi-relation environments. The approach preserved data privacy while achieving near-centralized performance.

Significance

This work addresses a critical challenge in knowledge graph question answering—performing multi-hop reasoning in a privacy-preserving, distributed setting. By integrating symbolic relation rules with embedding-based reasoning, it overcomes the limitations of traditional centralized systems and horizontal federated models. The framework enables organizations to collaborate on complex reasoning tasks without exposing raw data, opening new avenues for secure, scalable knowledge-driven applications across industries such as healthcare, finance, and enterprise data management. Its robustness and efficiency suggest broad applicability, paving the way for future integration with large language models and dynamic knowledge updates.

Technical Contribution

The paper introduces a novel federated framework that combines local graph enrichment, relation rule expansion, and entity anchoring, enabling multi-hop reasoning over vertically partitioned graphs. It innovatively fuses entity embeddings from multiple silos via concatenation, avoiding privacy leaks of triples or relation parameters. The topic entity anchoring mechanism grounds questions in the correct subgraph neighborhood without runtime cross-silo communication. The approach supports flexible relation subset partitioning, maintains high answer coverage, and achieves near-centralized performance, representing a significant step forward in federated symbolic reasoning.

Novelty

This is the first framework to support multi-hop question answering over vertically partitioned knowledge graphs with disjoint relation types, combining symbolic rule expansion, entity anchoring, and embedding fusion. Unlike prior federated KGE or distributed QA systems, it explicitly addresses the challenge of cross-silo reasoning in relation-disjoint environments, filling a major gap in the literature. Its unique integration of rule-based graph enhancement and neural embedding alignment sets it apart from existing methods.

Limitations

The reliance on predefined relation rules limits adaptability to dynamic or overlapping relation environments. Handling relation overlaps or evolving schemas requires further development.
The approach assumes static graphs and precomputed candidate sets, which may not scale efficiently to highly dynamic or large-scale graphs.
While privacy is preserved in data exchange, the embeddings may still leak some information, and formal privacy guarantees like differential privacy are not provided. Computational costs increase with the number of relations and entities.

Future Work

Future research will focus on automatic relation rule learning, dynamic schema adaptation, and integrating large pre-trained language models for enhanced question understanding. Exploring privacy-preserving guarantees like differential privacy and extending the framework to multi-modal knowledge graphs will broaden its applicability. Additionally, optimizing communication protocols and scalability for industrial-scale graphs remains a key direction.

AI Executive Summary

Knowledge graph question answering (KGQA) has become a vital tool for extracting structured information from vast, interconnected data sources. However, real-world applications often involve data distributed across multiple organizations, each holding partial views of the same entities but with disjoint relation types. Traditional centralized systems, while effective, pose privacy risks and are infeasible due to governance constraints. Existing federated approaches mainly address link prediction or simple queries in horizontally partitioned graphs, leaving multi-hop reasoning over vertically partitioned data largely unexplored.

This paper introduces FedV-KGQA, a novel framework designed to enable multi-hop reasoning over knowledge graphs that are partitioned by relation type across different organizations. The core idea is to combine local graph enrichment—using shared relation axioms like inverse and chain axioms—with entity embedding fusion and a topic entity anchoring mechanism. Each silo trains local knowledge graph embeddings (e.g., TransE, DistMult) without sharing raw triples, only transmitting entity embeddings to the server. The server fuses these embeddings via concatenation, projects questions using pre-trained transformers, and anchors questions to topic entities, enabling answer ranking without real-time cross-silo communication.

Experimental results on MetaQA, WebQSP, and PathQuestion datasets demonstrate that FedV-KGQA achieves performance close to centralized systems, supporting up to 3-hop reasoning with high answer coverage and robustness to embedding perturbations. The approach effectively balances privacy preservation, computational efficiency, and reasoning capability, marking a significant advance in federated symbolic reasoning. Its broad applicability across domains like healthcare, finance, and enterprise data underscores its potential to reshape distributed knowledge management.

Despite its strengths, the framework relies on static relation schemas and precomputed candidate sets, which may limit scalability and adaptability in dynamic environments. Future work aims to incorporate automatic rule learning, schema evolution, and integration with large language models, further enhancing its versatility and real-world impact.

Deep Dive

Plain Language Accessible to non-experts

想象你在一个大工厂工作，工厂里有很多不同的部门，每个部门负责不同的任务，比如一个负责装配，一个负责包装。每个部门都知道自己负责的部分，但不愿意把所有信息都告诉别人。现在，你想知道从原材料到成品的完整流程，但信息被分散在不同部门。你不能直接问所有部门，因为他们不想泄露秘密。于是，你用一些聪明的方法，比如用规则帮忙推断，或者用一个共同的线索（比如产品的编号）来锚定信息。这样，你就可以在不打扰别人的情况下，找到答案。这就像FedV-KGQA一样，多个“部门”合作推理，既保护隐私，又找到正确答案。

ELI14 Explained like you're 14

想象你在学校里，有很多不同的班级，每个班级都知道一些关于学生的秘密，但他们不想把全部信息都告诉别人。现在，你想知道哪个学生参加了学校的舞蹈比赛，但这个信息可能在不同班级里。你不能直接问所有班级，因为他们不想泄露秘密。于是，你用一些聪明的方法，比如用规则帮忙推断，或者用一个共同的线索（比如学生的名字）来锚定信息。这样，你就可以在不打扰别人的情况下，找到答案。这就像FedV-KGQA一样，多个“班级”合作推理，既保护隐私，又找到正确答案。

Glossary

Knowledge Graph (知识图谱)

一种用实体和关系表示知识的图形结构，支持推理和问答。

论文中用于描述存储事实的结构。

Knowledge Graph Embedding (知识图谱嵌入)

将实体和关系映射到低维向量空间，用于计算相似度和推理。

模型训练和路径推理的基础。

Vertical Federated Learning (垂直联邦学习)

不同机构共享实体空间但拥有不同关系或特征的分布式学习方式。

本文的主要场景。

Multi-Hop Question Answering (多跳问答)

通过多步推理连接多个事实，找到问题的答案。

论文的核心任务。

Relation Rule Expansion (关系规则扩展)

利用关系逆元和关系链规则，增强图结构，支持多跳推理。

提高多跳推理路径的完整性。

Open Questions Unanswered questions from this research

1 关系类型重叠或动态变化的知识图谱中，保持推理路径完整性仍未解决，现有方法多依赖静态规则，未来需探索自动规则学习与更新机制。
2 在大规模、多关系环境中，计算和通信成本仍偏高，需优化算法以适应实际应用。

Applications

Immediate Applications

企业知识平台

支持多机构在不泄露敏感数据的前提下，共同构建和推理知识图谱，提升业务协作效率。

医疗信息整合

实现不同医院间的患者信息推理，保护隐私同时支持复杂诊断。

Long-term Vision

行业智能问答

构建全面行业知识库，支持多机构协作的智能问答，推动行业数字化转型。

Abstract

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.

cs.AI