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

Md Saikat Islam Khan Bappy; Oshani Seneviratne

cs.AIAug 25, 2026

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

Authors: Md Saikat Islam Khan Bappy, Oshani Seneviratne

Organizations: Rensselaer Polytechnic Institute, Troy, NY 12180, USA

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.

Figures & tables

Figure 1: Motivating example of multi-hop KGQA in the vertical federated setting.

System	VFL	Multi-hop	KGE fusion	NL
EmbedKGQA [ 24 ]	✗	✓	✗	✓
UniKGQA [ 12 ]	✗	✓	✗	✓
RelChain [ 13 ]	✗	✓	✗	✓
RoG [ 18 ]	✗	✓	✗	✓
GMeLLo [ 6 ]	✗	✓	✗	✓
FedE [ 4 ]	✗	✗	✓	✗

Table 1: System comparison. VFL: vertical split, disjoint relations; Multi-hop: paths

\geq 2

; KGE fusion: representations fused across silos; NL: natural language questions.

Figure 2: Overview of the FedV-KGQA architecture.

Model	Scoring function $\phi(h,r,t)$	Space	$d_{k}$
TransE [ 2 ]	$-\\|\mathbf{h}+\mathbf{r}-\mathbf{t}\\|_{2}$	$\mathbb{R}^{d}$	$d$
DistMult [ 35 ]	$\sum_{j}h_{j}r_{j}t_{j}$	$\mathbb{R}^{d}$	$d$
RotatE [ 29 ]	$-\\|\mathbf{h}\circ e^{i\mathbf{r}}-\mathbf{t}\\|_{2}$	$\mathbb{C}^{d}$	$2d$
ComplEx [ 31 ]	$\operatorname{Re}(\langle\mathbf{h},\mathbf{r},\bar{\mathbf{t}}\rangle)$	$\mathbb{C}^{d}$	$2d$

Table 2: KGE scoring functions (

d_{k}=d

: real-valued;

d_{k}=2d

: complex-valued).

KGE	Encoder	Silo-3				Silo-5				Silo-7
		MRR	H@3	H@5	H@10	MRR	H@3	H@5	H@10	MRR	H@3	H@5	H@10
TransE	BERT	0.76	0.83	0.88	0.93	0.83	0.87	0.89	0.91	0.74	0.78	0.80	0.84
	DistilBERT	0.76	0.83	0.89	0.94	0.82	0.86	0.88	0.91	0.74	0.79	0.81	0.82
	RoBERTa	0.75	0.82	0.88	0.93	0.79	0.85	0.88	0.92	0.58	0.71	0.79	0.81
DistMult	BERT	0.76	0.82	0.87	0.93	0.83	0.86	0.89	0.92	0.69	0.77	0.79	0.82
	DistilBERT	0.76	0.82	0.88	0.93	0.83	0.87	0.89	0.92	0.70	0.77	0.79	0.82

Table 3: Results on the MetaQA dataset. The bold font denotes the best result.

KGE	Encoder	Silo-3				Silo-5				Silo-7
		MRR	H@3	H@5	H@10	MRR	H@3	H@5	H@10	MRR	H@3	H@5	H@10
TransE	BERT	0.54	0.59	0.64	0.71	0.52	0.57	0.62	0.69	0.51	0.56	0.62	0.67
	DistilBERT	0.53	0.59	0.64	0.73	0.52	0.57	0.64	0.71	0.50	0.56	0.62	0.69
	RoBERTa	0.43	0.49	0.54	0.61	0.41	0.47	0.53	0.60	0.39	0.44	0.50	0.57
DistMult	BERT	0.41	0.45	0.50	0.58	0.42	0.46	0.51	0.58	0.43	0.47	0.54	0.60
	DistilBERT	0.40	0.43	0.49	0.57	0.41	0.45	0.50	0.57	0.41	0.45	0.51	0.57

Table 4: Results on the WebQSP dataset. The bold font denotes the best result.

KGE	Encoder	Silo-3				Silo-5				Silo-7
		MRR	H@3	H@5	H@10	MRR	H@3	H@5	H@10	MRR	H@3	H@5	H@10
TransE	BERT	0.65	0.81	0.89	0.96	0.68	0.84	0.90	0.96	0.55	0.71	0.78	0.90
	DistilBERT	0.64	0.76	0.87	0.93	0.65	0.79	0.86	0.94	0.63	0.76	0.81	0.93
	RoBERTa	0.44	0.53	0.75	0.87	0.50	0.67	0.81	0.91	0.48	0.64	0.75	0.91
DistMult	BERT	0.64	0.80	0.85	0.93	0.63	0.81	0.88	0.94	0.62	0.79	0.87	0.91
	DistilBERT	0.66	0.79	0.88	0.94	0.63	0.76	0.87	0.92	0.63	0.77	0.84	0.93

Table 5: Results on the PathQuestion dataset. The bold font denotes the best result.

Dataset	Encoder + KGE	MRR		H@3		H@5		H@10
		2-Hop	3-Hop	2-Hop	3-Hop	2-Hop	3-Hop	2-Hop	3-Hop
MetaQA	BERT + TransE	0.76	0.74	0.83	0.82	0.88	0.87	0.93	0.93
	DistilBERT + ComplEx	0.76	0.72	0.81	0.80	0.87	0.86	0.92	0.91
	RoBERTa + TransE	0.75	0.73	0.82	0.81	0.88	0.87	0.93	0.92
PathQuestion	BERT + TransE	0.65	0.57	0.81	0.69	0.89	0.80	0.96	0.90
	DistilBERT + DistMult	0.66	0.51	0.79	0.61	0.88	0.69	0.94	0.81

Table 6: Comparative 2-hop and 3-hop performance on the 3-silo partition.

Figure 3: Ablation study across three datasets (deltas from FedV-KGQA in parentheses).

Method	MetaQA				PathQuestion				WebQSP
Method	MRR	H@3	H@5	H@10	MRR	H@3	H@5	H@10	MRR	H@3	H@5	H@10
Adapted EmbedKGQA [ 24 ]	0.26	0.27	0.30	0.35	0.32	0.36	0.44	0.51	0.30	0.34	0.37	0.42
Adapted FL-KG-QA [ 10 ]	0.45	0.47	0.48	0.48	0.31	0.34	0.36	0.42	0.29	0.32	0.34	0.37
Adapted FedE [ 4 ]	0.70	0.77	0.83	0.88	0.58	0.68	0.78	0.87	0.46	0.51	0.57	0.62
Adapted RelChain [ 13 ]	0.71	0.78	0.83	0.88	0.57	0.65	0.74	0.88	0.42	0.48	0.53	0.60
FedV-KGQA (Ours)	0.76	0.83	0.88	0.93	0.65	0.81	0.89	0.96	0.54	0.59	0.64	0.71

Table 7: Comparison with adapted baselines. Best results are shown in bold .

Figure 4: FedV-KGQA robustness to Gaussian noise (BERT+TransE, five silos). $\sigma=0$ denotes the no-noise baseline.

Figure 5: Communication cost in GB to reach H@3 targets across encoders (TransE, three silos). A dash (—) indicates the target was not reached.

Knowledge Graph Question Answering

Knowledge Graph Question Answering answers natural-language questions by finding and reasoning over entities and relations in structured knowledge graphs.

See all 165 papers

FedV-KGQA in Practice: Design Lessons and an Interactive Prototype
Sep 12, 2026Md Saikat Islam Khan Bappy, Oshani SeneviratneKnowledge Graph Question Answering Multi-Hop QA
Rensselaer Polytechnic Institute, Troy, NY 12180, USA
Cost Characterization of Vertically Partitioned Federated Knowledge Graphs
Sep 12, 2026Md Saikat Islam Khan Bappy, Oshani SeneviratneKnowledge Graphs
Rensselaer Polytechnic Institute, Troy, NY 12180, USA
Compositional Chain-of-Relations for Faithful Knowledge Graph Question Answering with Large Language Models
Aug 24, 2026Chenhui Liu, Jianpeng Zhou, Jiahai WangKnowledge Graph Question Answering LLM Reasoning with Graphs
Sun Yat-sen University, Guangzhou 510006, China

Sep 12, 2026J/K move · Enter open · S save

FedV-KGQA in Practice: Design Lessons and an Interactive Prototype

Md Saikat Islam Khan Bappy, Oshani Seneviratne

Rensselaer Polytechnic Institute, Troy, NY 12180, USA

Knowledge Graph Question Answering Multi-Hop QA Federated Learning KG Embedding

Open paper AskPDF arXiv

Knowledge graph question answering usually assumes that one system can reach the whole graph. In practice, facts are often held by organizations that share entity identifiers but own disjoint relation types, so no single party sees a complete reasoning chain. This poster presents the empirical findings of FedV-KGQA on multi-hop question answering over such vertically partitioned graphs. Each silo enriches its local graph and trains a knowledge graph embedding on its own triples. A server then concatenates the silo-specific entity views, anchors the projected question at the topic entity, and ranks candidates by similarity. Raw triples and relation embeddings never leave a silo. Comparing the FedV-KGQA experiments with one another yields three results. First, federated fusion recovers most of the centralized accuracy, while a single silo recovers little. Second, anchoring and enrichment matter more than the choice of embedding model. Third, the cheapest encoder depends on the target accuracy rather than on parameter count. This poster paper contributes that cross-experiment comparison, four design lessons drawn from it, and an interactive prototype that runs real inference and traces the full pipeline, per question, on released checkpoints.

Abstract

Knowledge Graph Question Answering

Explore similar work

FedV-KGQA in Practice: Design Lessons and an Interactive Prototype

Cost Characterization of Vertically Partitioned Federated Knowledge Graphs

Compositional Chain-of-Relations for Faithful Knowledge Graph Question Answering with Large Language Models