Towards Training and Inference of Graph Neural Networks in Graph Databases

Victor Pekkari; Harry Wik; Alfred Clemedtson; Borun Shi; Xuan-Son Vu

cs.LGOct 13, 2026

Towards Training and Inference of Graph Neural Networks in Graph Databases

Authors: Victor Pekkari, Harry Wik, Alfred Clemedtson, Borun Shi, Xuan-Son Vu

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

Abstract

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.

Figures & tables

Figure 1. Illustration of computations that happen in the training process and the graph database when processing a batch. Note that the training process can live in the same or different machine to the graph DB.

	Cora	CoAuthor	arXiv	products	papers100M
# nodes	2,708	34,493	169,343	2,449,029	111,059,956
# edges	10,556	495,924	1,166,243	61,859,140	1,615,685,872
Avg degree	3.90	14.38	6.89	25.26	14.55
Training nodes	140	20,696	90,941	196,615	1,207,179
Validation nodes	500	3,449	29,799	39,323	125,268
Test nodes	1,000	10,348	48,603	2,213,091	214,338

Table 1. Dataset statistics and training configuration.

Figure 2. Validation accuracy on Cora for GraphSAGE and GraphSAINT sampling implemented in PyG, Cypher and UDP. Each line shows the mean validation accuracy over 10 runs, and the shaded region shows the 95% interval.

Figure 3. ogbn-papers100M , GraphSAGE-sampling. Breakdown of time spent in training. baseline_neo4j represents Cypher implementation. Using UDP drastically reduces both data transfer and sampling execution time. Using UDF speeds up baseline_neo4j by 2̃.5 times. Using pre-aggregation does not provide improvement in latency for the when feature dimension is low in this case.

Figure 4. Per-operator execution-time breakdown of the Cypher GraphSAGE sampler on ogbn-papers100M . Graph traversal ( Expand(All) ) accounts for roughly $14.6$ ms, less than the combined cost of Projection and OrderedAggregation , which only build and sort intermediate lists rather than touch the graph.

Figure 5. Database hits vs. rows produced per operator for the GraphSAGE sampling-query on ogbn-papers100M (log-log, y-axis starts at 100). The Expand operator produces over 10,000 rows, well above the necessary $128\cdot(1+10+10\cdot 5)=7808$ nodes, for fanout [10, 5] and batch_size=128 , showing that the database fetches raw relationships later discarded during sampling.

Figure 6. GraphSAGE-sampling. Pre-aggregation reduces data sent from the DB to training process between ca. 40% - 60% for Fanout [10, 5] .

Figure 7. CoAuthor Physics . GraphSAGE-sampling. Having the highest feature dimensionality, this dataset benefits most from pre-aggregation, as the added DB-side compute and querying cost is outweighed by lower Network + driver recv time from sending less data.

Panel of Figure 8 — Figure 8. Benchmark of inference in databases.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9. Representing embedding properties as byte[] reduces the time taken by driver by more than half, compared to float64[] format, due to the large reduction in the packstream decode operation in the Bolt client driver.

Figure 10. ogbn-products . Breakdown of time spent in training.

Explore similar work

Taurus: Accelerating Out-of-Core Graph Neural Network Inference on Billion-Scale Graphs
Jul 19, 2026Pranjal Naman, Yogesh SimmhanEfficient Neural Network Inference Graph Neural Networks
Department of Computational and Data Sciences (CDS), Indian Institute of Science (IISc), Bangalore 560012 India
SNI-GNN: SmartNIC-Assisted Full-Graph GNN Training with In-Network Embedding Prediction
Aug 6, 2026Guofan Yu, Sitian Chen, Zhenheng Tang +2Graph Neural Networks Communication-Efficient Distributed Training
Hong Kong Baptist University · Hong Kong University of Science and Technology · Hong Kong University of Science and Technology (Guangzhou)
An LLM-Guided Query-Aware Inference System for GNN Models on Large Knowledge Graphs
Mar 4, 2026Waleed Afandi, Hussein Abdallah, Ashraf Aboulnaga +1Efficient Neural Network Inference Graph Neural Networks
Concordia University · University of Texas at Arlington

Jul 19, 2026J/K move · Enter open · S save

Taurus: Accelerating Out-of-Core Graph Neural Network Inference on Billion-Scale Graphs

Pranjal Naman, Yogesh Simmhan

Department of Computational and Data Sciences (CDS), Indian Institute of Science (IISc), Bangalore 560012 India

Efficient Neural Network Inference Graph Neural Networks Memory-Efficient Inference

Open paper AskPDF arXiv

Graph Neural Network (GNN) inference on billion-scale graphs is challenging due to the large memory footprint of features and embeddings and high disk I/O costs in out-of-core settings. Existing distributed GNN systems incur high communication times and infrastructure costs, while disk-based GNN systems are primarily tailored to training and experience massive wasted reads during inference on the entire graph. We present Taurus, a single-machine system for GNN inference on graphs that do not fit in RAM, supporting both exact full-graph inference and fanout-sampled inference. To avoid random and repeated feature gathers, Taurus reformulates layer-wise inference as source-centric broadcasts over sequential SSD scans, backed by a pipelined GPU-CPU-SSD hierarchy, topology-aware reordering, pending-message eviction, and a GPU-resident store for high-degree vertices. It further uses non-buffered sequential reads and GPU-backed writes to reduce page-cache pollution, host-memory pressure, and write overheads. On out-of-core graphs with up to

269M

vertices,

4B

edges, and

514

GiB of features, Taurus outperforms the strongest layer-wise baseline, DGI, by

7

25\times

, and vertex-wise baselines by

40

140\times

Abstract

Explore similar work

Taurus: Accelerating Out-of-Core Graph Neural Network Inference on Billion-Scale Graphs

SNI-GNN: SmartNIC-Assisted Full-Graph GNN Training with In-Network Embedding Prediction

An LLM-Guided Query-Aware Inference System for GNN Models on Large Knowledge Graphs