2026-10-01T22:37:51Z https://drops.dagstuhl.de/oai

oai:drops-oai.dagstuhl.de:28088 2026-09-30T14:08:58Z ddc:004 open_access

A Hybrid Vector-Symbolic Data Model for Multimodal Knowledge Graphs Ruosch, Florian Rossetto, Luca Multimodal Knowledge Graphs Graph Store Multimodal Media Segmentation In this article, we address the architectural and representational “schism” between symbolic knowledge graphs and sub-symbolic multimodal data. By formalizing the MediaGraph Data Model, we demonstrate that high-dimensional vectors and multimedia segments can be integrated as first-class citizens within a unified graph structure. This shift from treating vectors as opaque literals to queryable nodes enables “early-binding” query execution, allowing the SPARQL engine to jointly optimize over both structural relationships and perceptual similarity. Our implementation, MeGraS, provides a scalable engine that mitigates the long-standing N+1 query problem through batched execution. The experimental evaluation on both real-world lifelogging data and a synthetic dataset of 10⁷ triples confirms that our hybrid approach maintains high throughput and retrieval efficiency, even as the complexity of multimodal queries increases. Furthermore, the integration of the Unified Multimedia Segmentation Model ensures that media granularity is preserved, making complex spatiotemporal and content-aware querying possible within a single query framework. Ultimately, MeGraS provides a foundational infrastructure for the next generation of multimodal intelligent systems. By unifying the symbolic and sub-symbolic domains, we provide a pathway toward more expressive and efficient retrieval strategies that can keep pace with the explosion of heterogeneous digital media. Schloss Dagstuhl – Leibniz-Zentrum für Informatik Florian Ruosch and Luca Rossetto 2026 Is Part Of TGDK, Volume 4, Issue 3 (2026). Transactions on Graph Data and Knowledge, Volume 4, Issue 3 Article Text doc-type:ResearchArticle publishedVersion application/pdf doi:10.4230/TGDK.4.3.3 urn:nbn:de:0030-drops-280880 https://drops.dagstuhl.de/entities/document/10.4230/TGDK.4.3.3 eng https://creativecommons.org/licenses/by/4.0/legalcode