A Three-stage Neuro-symbolic Recommendation Pipeline for Cultural Heritage Knowledge Graphs

2026-02-23

AI Summary

This study addresses the challenge of generating accurate and interpretable recommendations in digital cultural heritage contexts, where heterogeneous data exhibit complex semantic relationships. To this end, it proposes the first neuro-symbolic recommendation framework specifically designed for cultural heritage knowledge graphs. The approach integrates knowledge graph embeddings—employing models such as TransE, ComplEx, ConvE, and CompGCN—with HNSW-based approximate nearest neighbor search and SPARQL-powered semantic filtering to construct a three-stage explainable recommendation pipeline. This architecture effectively balances recommendation quality and interpretability, even under conditions of sparse and heterogeneous data. Experimental evaluation on the JUHMP knowledge graph, comprising 3.2 million RDF triples, demonstrates that the generated recommendations are both practically useful and semantically sound, as validated by domain experts.

Technology Category

• Data Mining & Knowledge Management: Recommender Systems
• Machine Learning: Neuro-Symbolic Learning
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Semantics and Knowledge: Scalable techniques for the creation, curation, publication, maintenance, and consumption of large, Web-based, structured, reusable, knowledge graphs and ontologies
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

The growing volume of digital cultural heritage resources highlights the need for advanced recommendation methods capable of interpreting semantic relationships between heterogeneous data entities. This paper presents a complete methodology for implementing a hybrid recommendation pipeline integrating knowledge-graph embeddings, approximate nearest-neighbour search, and SPARQL-driven semantic filtering. The work is evaluated on the JUHMP (Jagiellonian University Heritage Metadata Portal) knowledge graph developed within the CHExRISH project, which at the time of experimentation contained ${\approx}3.2$M RDF triples describing people, events, objects, and historical relations affiliated with the Jagiellonian University (Kraków, PL). We evaluate four embedding families (TransE, ComplEx, ConvE, CompGCN) and perform hyperparameter selection for ComplEx and HNSW. Then, we present and evaluate the final three-stage neuro-symbolic recommender. Despite sparse and heterogeneous metadata, the approach produces useful and explainable recommendations, which were also proven with expert evaluation.

Problem

• cultural heritage
• knowledge graph
• recommendation system
• semantic relationships
• heterogeneous data

Innovation

• neuro-symbolic recommendation
• knowledge graph embeddings
• approximate nearest neighbour search
• SPARQL-based semantic filtering
• cultural heritage