Research map / Graph learning / Knowledge and temporal graphs

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

NeurIPS 2026 · Graph learning › Knowledge and temporal graphs · approach: Logical query answering and rule reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NGDB-Zoo, an operator-level batching and dynamic scheduling training framework for neural graph databases, plus decoupled GPU-resident integration of pretrained text encoder embeddings, reporting 1.8x-6.8x throughput gains on complex query answering benchmarks.

Problem

Training neural graph databases for complex logical query answering is hampered by query-level batching that fragments GPU kernels on diverse query structures, and by the I/O and memory costs of integrating high-dimensional pretrained text encoder semantics into structure-only entity embeddings.

Approach

Decompose logical queries into DAGs of atomic operators (projection, intersection, union, negation) pooled across queries by type and scheduled with a max-fillness policy for cross-query fusion; pre-compute frozen text-encoder entity embeddings offline and cache them GPU-resident so the training loop is inference-free.

Key results

• On FB15k with BetaE, NGDB-Zoo reaches 4,477 queries/sec, a 7.0x speedup over SQE, with MRR 43.04 vs 40.82 for SQE.
• Adding semantic optimizations with Qwen3-Embedding gives 5x-7x throughput gains and an average MRR increase from 26.32 to 31.06 across FB15k, FB15k-237 and NELL995 variants.
• Single-hop KG completion on Freebase with 1 GPU takes 628 s per epoch versus 760 s for SMORE and 3060 s for PBG (V100 numbers from prior work).

Surprising findings

• Reported GPU memory for NGDB-Zoo is higher than SMORE on several small-KG settings (e.g., BetaE on FB15k: 12.95 GB vs 1.16 GB), a tradeoff the authors attribute to GPU-resident caching.

Limitations

• Text encoder is frozen, decoupling semantic and structural learning and limiting adaptation to graph topology.
• Large 4M-node graphs rely on CPU offloading, making throughput dependent on PCIe bandwidth.
• Some memory usage is higher than SMORE due to GPU-resident caching.

Experimental setup

Datasets: FB15k FB15k-237 NELL995 FB400k ogbl-wikikg2 ATLAS-Wiki-Triple-4M

Benchmarks: FB15k FB15k-237 NELL995 FB400k ogbl-wikikg2 ATLAS-Wiki-Triple-4M

Metrics: MRR training throughput (queries/sec) GPU memory (GB) Hit@10

Baselines: KGReasoning SMORE SQE Marius PBG DGL-KE GQE Q2B Q2P FuzzQE

Builds on: BetaE Atom SMORE SQE

Closest papers by meaning

• Inductive Logical Query Answering in Knowledge Graphs (NeurIPS 2022 · similarity 0.86)
• Rethinking Complex Queries on Knowledge Graphs with Neural Link Predictors (ICLR 2024 · similarity 0.85)
• Talk like a Graph: Encoding Graphs for Large Language Models (ICLR 2024 · similarity 0.85)
• Griffin: Towards a Graph-Centric Relational Database Foundation Model (ICML 2025 · similarity 0.85)
• RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases (NeurIPS 2025 · similarity 0.85)
• Finding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language Models (ICML 2026 · similarity 0.84)