What question did this study set out to answer?

This study aims to enhance knowledge graph retrieval through a hybrid method integrating multiple retrieval techniques.

May 13, 2026Open Access

Topology-Aware Binary SDM for Knowledge Graph Retrieval: A Multi-Architecture Empirical Study with Neural Baseline and Quantum Walk Analysis

Key Points

This study aims to enhance knowledge graph retrieval through a hybrid method integrating multiple retrieval techniques.
Introduced Topology-Aware Sparse Distributed Memory (TA-SDM) using SimHash and graph neighbor aggregation.
Evaluated retrieval performance on a 392-node heterogeneous graph with continuous-time quantum walks.
Conducted experiments to compare TA-SDM with traditional methods regarding storage and accuracy.
TA-SDM achieved a mean reciprocal rank (MRR) of 0.914 ± 0.038 with Recall@5 of 0.676 ± 0.037.
Demonstrated a 3.45x improvement over SimHash and 2.13x over a neural embedding baseline, both with p < 0.001.
The method used 48x less storage per node compared to traditional approaches.

Abstract

Knowledge graph retrieval — the task of finding relevant nodes near a query in a typed, heterogeneous graph — is a primitive underlying many information systems. State-of-the-art methods (dense neural embeddings indexed with GPU-accelerated approximate nearest neighbor search) impose substantial storage, API, and hardware costs. In this paper we introduce and empirically evaluate a hybrid retrieval method that combines three previously disconnected techniques: (1) SimHash content addressing with a weighted majority-vote aggregation of 1-hop graph neighbor signatures, producing 256-bit binary node addresses we call Topology-Aware Sparse Distributed Memory (TA-SDM); and (2) classical simulation of continuous-time quantum walks (CTQW) on BFS-extracted subgraphs. On a 392-node heterogeneous typed knowledge graph, TA-SDM achieves MRR of 0.914 ± 0.038 (mean over 10 seeds; 95% CI 0.891, 0.937) with Recall@5 of 0.676 ± 0.037, a 3.45x improvement over content-only SimHash (p 0.05). The method requires no neural training, no GPU, no embedding API, and no quantum hardware; the complete implementation uses only the Python standard library and hardware POPCNT instructions. We validate reproducibility across three CPU generations spanning thirteen years (Intel Sandy Bridge 2011, Tiger Lake 2020, and Arrow Lake 2024): output is bit-exact identical on all three machines despite up to 4.5x throughput differences, confirming that retrieval quality is a property of the algorithm rather than of hardware. We further report a structured literature review of 47 adjacent prior works from five distinct research traditions (SDM, hyperdimensional computing, locality-sensitive hashing, graph neural networks, and continuous-time quantum walks), identifying the specific combinatorial gap our construction fills. Finally, we document four negative results — on multi-hop aggregation, hippocampal place-cell encoding, reservoir computing, and compressed sensing — that constrain the design space.

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper