Combining different types of ontology alignment methods through voting-based ensemble learning reliably improves results, with the best approach depending on whether you prioritize precision (mix different paradigms) or overall F1-score (use multiple LLMs).
This paper presents OntoAligner-Ensemble, a framework that combines predictions from different ontology alignment methods (string-based, knowledge graph embeddings, and LLM-based) using voting and selection strategies. Testing across biomedical and other domains shows that mixing diverse alignment approaches consistently improves precision-recall balance and often beats individual methods.