NGDB-Zoo: Towards Efficient and Scalable Neural Graph Databases Training

Zhongwei Xie; Jiaxin Bai; Shujie Liu; Haoyu Huang; Yufei Li; Yisen Gao; Hong Ting Tsang; Yangqiu Song

cs.LGFeb 25, 2026

NGDB-Zoo: Towards Efficient and Scalable Neural Graph Databases Training

Authors: Zhongwei Xie, Jiaxin Bai, Shujie Liu, Haoyu Huang, Yufei Li, Yisen Gao, Hong Ting Tsang, Yangqiu Song

Organizations: The Hong Kong University of Science and Technology · Microsoft Research Asia

Abstract

Neural Graph Databases (NGDBs) facilitate complex logical reasoning over incomplete knowledge structures, yet their training efficiency and expressivity are constrained by rigid query-level batching and structure-exclusive embeddings. We present NGDB-Zoo, a unified framework that resolves these bottlenecks by synergizing operator-level training with semantic augmentation. By decoupling logical operators from query topologies, NGDB-Zoo transforms the training loop into a dynamically scheduled data-flow execution, enabling multi-stream parallelism and achieving a $1.8\times$ - $6.8\times$ throughput compared to baselines. Furthermore, we formalize a decoupled architecture to integrate high-dimensional semantic priors from Pre-trained Text Encoders (PTEs) without triggering I/O stalls or memory overflows. Extensive evaluations on six benchmarks, including massive graphs like ogbl-wikikg2 and ATLAS-Wiki, demonstrate that NGDB-Zoo maintains high GPU utilization across diverse logical patterns and significantly mitigates representation friction in hybrid neuro-symbolic reasoning.

Figures & tables

Figure 1: Task Illustration. Query embedding methods aim to answer multihop logical queries (A) by avoiding explicit knowledge graph traversal and executing the query directly in the embedding space by following the query computation plan (B). Operators are executed in the embedding space (C).

Figure 2: Evolution toward High-Throughput Asynchronous Pipelining. Training methods. (a) represents the naive training scheme; (b) shows the pre-sampling and pre-fetching optimization; (c) NGDB-Zoo’s operator-level approach breaks these constraints to enable fully overlapping stages, creating a dense execution stream that maximizes hardware saturation.

Figure 3: Overcoming Topological Rigidity. (Left): Query-level batching incurs high fragmentation overheads by segregating distinct query structures. (Right): NGDB-Zoo aggregates atomic operators into homogeneous Operator Pools for unified execution, eliminating structural constraints and saturating GPU cores.

Figure 4: Max-Fillness Dynamic Scheduling. By prioritizing operator pools with the highest workload, the scheduler maximizes GPU utils. This dynamic selection effectively unblocks dependent nodes, ensuring a continuous, high-throughput training stream.

Figure 5: Vectorized execution for Intersection/Union. By grouping operations into cardinality-based Equivalence Classes, we eliminate tensor misalignment. This enables perfectly aligned Vectorized Execution , replacing slow loops with dense matrix operations.

Figure 6: Illustration of the decoupled semantic integration strategy. By relegating the heavy text encoder (PTE) to an offline pre-computation phase and caching embeddings in GPU memory, NGDB-Zoo achieves minimal-overhead semantic fusion during training.

Dataset	Model	MRR (%)	TPut ( $\uparrow$ )	Mem (GB $\downarrow$ )
FB400k	GQE	35.84	24.68	7.5
	Q2B	52.33	21.55	11.0
	BetaE	50.40	19.97	14.0
ogbl-wikikg2	GQE	32.88	23.92	8.0
	Q2B	42.01	20.75	11.0
	BetaE	44.54	19.65	14.0

Table 1: Scalability and Predictive Performance on Massive Knowledge Graphs. We report the filtered MRR (%), training throughput (

10^{3}

Queries/sec), and peak GPU memory occupancy (GB) across three production-scale datasets. NGDB-Zoo enables high throughput and robust reasoning across diverse paradigms on production-scale graphs with millions of entities.

System	Epoch Time (s)
System	1-GPU	2-GPU	4-GPU	8-GPU
Marius ( Mohoney et al., 2021 ) *	727	-	-	-
PBG ( Lerer et al., 2019 ) *	3060	1400	515	419
SMORE ( Ren et al., 2022 ) *	760	411	224	121
NGDB-Zoo (Ours)	628	322	181	94

Table 2: Runtime performance on Freebase KG. *Results taken from Mohoney et al. (2021) and Ren et al. (2022) with the same V100 GPUs.

		MRR (%)				Training queries (Queries/Sec)				GPU Memory (GB)
Dataset	Model	KGR	SMORE	SQE	NGDB-Zoo	KGR	SMORE	SQE	NGDB-Zoo	KGR	SMORE	SQE	NGDB-Zoo
FB15k	BetaE	41.60	40.39	40.82	43.04	291	2808	636	4477	2.68	1.16	8.05	12.95
	Q2B	38.00	41.54	41.08	42.85	855	3588	343	4086	1.32	0.96	6.01	5.07
	GQE	28.00	30.60	29.21	30.39	1395	3770	4598	6271	0.76	1.09	8.03	11.33
	Q2P	-	-	44.30	46.90	-	-	832	1940	-	-	4.76	11.41
	FuzzQE	-	-	42.17	42.33	-	-	720	2973	-	-	20.14	8.45

Table 3: Performance comparison of NGDB-Zoo vs. KGReasoning, SMORE and SQE.

Figure 7: Multi-GPU Throughput Scalability. Training throughput increases near-linearly with the number of GPUs across ogbl-wikikg2 and ATLAS-Wiki-Triple-4M. Units are in # Queries (x1K/sec).

Figure 8: Impact of Decoupled Semantic Integration. We compare the standard joint training (Solid) against our GPU-Resident Decoupled strategy (Hatched //) across three datasets and models. (a) Predictive Performance: Our method consistently improves MRR than without semantic priors by leveraging semantic priors. (b) Throughput: By removing the heavy Text Encoder from the training loop, we achieve a 5x-7x speedup . (c) Memory: Despite caching embeddings, peak memory is reduced because the encoder is unloaded. Note: Higher is better for (a)&(b); lower is better for (c).

Figure 9: Performance comparison with and without Adaptive Sampling. Adaptive Sampling consistently improves MRR over all models and datasets.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Entities	Relations	Train Edges	Valid Edges	Test Edges	Total Edges
FB15k	14,951	1,345	483,142	50,000	59,071	592,213
FB15k-237	14,505	237	272,115	17,526	20,438	310,079
NELL995	63,361	200	114,213	14,324	14,267	142,804
FB400k	409,829	918	1,075,837	537,917	537,917	2,151,671
ogbl-wikikg2	2,500,604	535	16,109,182	429,456	598,543	17,137,181
ATLAS-Wiki-Triple-4M	4,035,238	512,064	23,040,868	2,880,108	2,880,110	28,801,086

Appendix

Table 4: KG statistics with number of entities, relations, training, validation and test edges.

Parameter	Value
Batch Size	$512$ queries
Optimizer	Adam (lr $10^{-4}$ )
Margin ( $\gamma$ )	$12.0$

Appendix

Table 5: Hyperparameter settings and model configurations.

Operator	Baseline (ms)	Batched (ms)	Speedup
EmbedE	2.3	0.8	$2.88\times$
Project	15.7	4.2	$3.74\times$
Intersect	78.5	6.0	13.11 $\times$
Union	62.3	5.1	$12.22\times$

Appendix

Table 6: Comparison of execution time per operator: Baseline vs. Batched.

Dataset	Metrics	2in	3in	inp	pin	pni	avg
FB15k	MRR	13.00	14.97	9.17	6.11	11.88	11.03
FB15k	Hit@10	26.36	30.56	18.53	11.97	23.74	22.23
FB15k-237	MRR	3.96	6.95	6.52	3.97	2.96	4.87
FB15k-237	Hit@10	7.75	14.19	13.45	7.45	5.08	9.584
NELL995	MRR	4.06	6.65	8.03	3.25	2.92	4.98
NELL995	Hit@10	7.84	14.88	16.05	5.61	5.31	9.94

Appendix

Table 7: Performance (%) of BetaE on queries with negation.

		MRR (%)		Training queries (Queries/Sec)		GPU Memory (GB)
Dataset	Model	NGDB-Zoo	NGDB-Zoo + Optimizations	NGDB-Zoo	NGDB-Zoo + Optimizations	NGDB-Zoo	NGDB-Zoo + Optimizations
FB15k	BetaE +Qwen-EMB	42.04	46.92	209	1516	14.23	13.11
	Q2B +Qwen-EMB	41.85	50.86	348	2014	6.39	5.18
	GQE +Qwen-EMB	29.39	35.91	558	2254	12.61	11.49
	BetaE +BGE-Base-En-v1.5	42.04	45.37	187	1402	14.40	13.21
	Q2B +BGE-Base-En-v1.5	41.85	47.54	340	1886	6.55	5.23

Appendix

Table 8: Performance comparison with and without Semantic Entity Representations optimizations across three datasets (FB15k, FB15k-237, NELL995) and three models (BetaE, Q2B, GQE). Solid bars show baseline performance, hatched bars (//) show optimized performance. Higher is better for MRR and throughput; lower is better for memory. The optimizations consistently improve MRR across all models and datasets, dramatically increase training throughput, and modestly reduce memory consumption.

Explore similar work

Towards Training and Inference of Graph Neural Networks in Graph Databases
Oct 13, 2026Victor Pekkari, Harry Wik, Alfred Clemedtson +2Graph Neural Networks Efficient Neural Network Inference
Lund University, Faculty of Engineering, Lund, Sweden · Neo4j, London, England
NGDBench: Towards Neural Graph Data Management
Feb 27, 2026Yufei Li, Yisen Gao, Jiaxuan Xiong +6Graph Retrieval-Augmented Generation Text-to-Cypher
HKUST · Beijing Institute of Technology · Hong Kong Baptist University +1
NSFL: A Post-Training Neuro-Symbolic Fuzzy Logic Framework for Boolean Operators in Neural Embeddings
Apr 12, 2026Vladi Vexler, Ofer Idan, Gil Lederman +1Information Retrieval Neuro-Symbolic Reasoning
Huawei Tel-Aviv Research Center

Oct 13, 2026J/K move · Enter open · S save

Towards Training and Inference of Graph Neural Networks in Graph Databases

Victor Pekkari, Harry Wik, Alfred Clemedtson, Borun Shi+1

Lund University, Faculty of Engineering, Lund, Sweden · Neo4j, London, England

Graph Neural Networks Efficient Neural Network Inference

Open paper AskPDF arXiv

Graph Neural Networks on large graphs commonly rely on in-memory graph storage with ad hoc data processing pipelines for both training and inference. Graph databases provide scalable storage and efficient traversal for graphs that can evolve over time, but their applicability in GNN training and inference are rarely studied. We show that graph DBs can be used in GNN training and inference directly, executing sampling, feature retrieval, and even the full forward pass. We show different GNN training pipelines can perform on-the-fly sampling and retrieval in graph DBs using Cypher. We further provide optimisations that speed up the training process by ~2.5x while maintaining the same accuracy. We illustrate that fully in-DB inference is possible, independent of how the GNNs are trained. Our inference setup can be >10x faster than cold loading the graph into RAM with no extra memory overhead on real-world graphs. Our work lays the foundation for in-graph-DB GNN training and inference.

Abstract

Explore similar work

Towards Training and Inference of Graph Neural Networks in Graph Databases

NGDBench: Towards Neural Graph Data Management

NSFL: A Post-Training Neuro-Symbolic Fuzzy Logic Framework for Boolean Operators in Neural Embeddings