Integrating user-informed predicate weighting into knowledge graph embedding workflows significantly improves the semantic coherence of biomedical clustering, according to a study published in the Journal of Biomedical Semantics. The research introduces LISE1, an updated system designed to bridge the gap between mathematical representation and human domain relevance. By allowing user-derived signals to actively adjust the underlying embedding generation, the system steers automated clustering toward biologically meaningful groupings rather than relying solely on raw numerical proximity.
Knowledge Graph Embeddings have become an established tool across biomedical informatics, serving essential roles in automated clustering, similarity assessment, and biomedical knowledge discovery. Despite demonstrating robust capabilities in tasks such as link prediction, conventional embedding models encounter a persistent limitation: spatial proximity within dense vector spaces does not consistently correspond to meaningful semantic similarity. In specialized biomedical contexts, resources mapped close to one another computationally often fail to share coherent biological characteristics.
To address these alignment difficulties, researchers previously developed the Logic-based Interactive Similarity Explainer, or LISE, which exposed shared semantic properties among clustered Resource Description Framework entities and collected user evaluations regarding cluster coherence. However, earlier versions suffered from practical constraints in verbalization granularity and anaphora resolution due to a rigid template-based structure. In LISE1, the authors replaced the template system with a Gemini 2.5 Flash language module capable of generating fluent, path-level explanations of logical Common Subsumers, establishing reliable sentence-level evaluations for human users.
To determine whether explicit human preferences could directly guide graph representations, the investigators developed and evaluated two distinct sampling techniques within the pyRDF2Vec framework. The experimental setup simulated user feedback across 1,280 semantic triples sourced from the DrugBank database. The team tested Predicate Relevance Weight against Predicate-Object Relevance Weight to examine whether weighting broad relation types or specific relation-target pairs offered superior control over vector generation and downstream grouping.
The weighting system learned user relevance scores through a trained random forest regressor. When applying learned weights, predicate-level weighting successfully increased the proportion of user-favored predicates inside the most cohesive clusters, with every examined predicate displaying either a positive or neutral deviation. In contrast, the more granular predicate-object weighting exhibited limited sensitivity, leaving the frequency of most examined relation-object pairs largely unchanged regardless of the assigned weight values. Average deviation metrics confirmed that adjusting predicate weights alone redirected clustering far more effectively toward relevant biomedical properties.
The computational findings were subsequently corroborated through user evaluations conducted with 23 human participants. The participants reviewed clusters refined under the different weighting configurations and reported noticeably higher satisfaction when evaluating clusters organized through predicate-level weighting strategies. Overall user scoring improved following the interactive refinement process, confirming that embedding generation can be successfully steered by user-informed relevance feedback.
Ultimately, the researchers concluded that user-informed predicate weighting provides a viable and effective mechanism for enhancing embedding-based clustering in biomedical graphs. Because predicate-object adjustments offered minimal tangible benefit, tuning graph embeddings at the predicate level represents the more effective strategy for aligning automated clustering with biomedical domain semantics.
Subject of Research: Semantic tuning of biomedical knowledge graph embeddings for clustering
Article Title: Tuning knowledge graph embeddings in clustering with LISE1
Article References: Schena, V., Colucci, S., Donini, F. M., Scioscia, F., & Di Sciascio, E. (2026). Tuning knowledge graph embeddings in clustering with LISE1. Journal of Biomedical Semantics. https://doi.org/10.1186/s13326-026-00367-4
Image Credits: AI Generated
DOI: 10.1186/s13326-026-00367-4
Keywords: Knowledge graph embeddings, Biomedical informatics, LISE1, Semantic clustering, Gemini 2.5 Flash, Common Subsumers, DrugBank, pyRDF2Vec, Predicate Relevance Weight, Predicate-Object Relevance Weight, Random forest regressor, RDF resources, Cluster coherence, Similarity assessment, Knowledge discovery, Natural language explanation, Path-level explanations
Cite Scienmag News
Reid Dalton. (November 16, 2026). Predicate Weighting Improves Semantic Clustering in Biomedical Knowledge Graphs. Scienmag. https://scienmag.com/predicate-weighting-improves-semantic-clustering-in-biomedical-knowledge-graphs/
Reid Dalton. "Predicate Weighting Improves Semantic Clustering in Biomedical Knowledge Graphs." Scienmag, 16 November 2026, https://scienmag.com/predicate-weighting-improves-semantic-clustering-in-biomedical-knowledge-graphs/. Accessed 16 November 2026.
Reid Dalton. "Predicate Weighting Improves Semantic Clustering in Biomedical Knowledge Graphs." Scienmag. November 16, 2026. https://scienmag.com/predicate-weighting-improves-semantic-clustering-in-biomedical-knowledge-graphs/








