OntoAligner-Ensemble: Voting-Based Fusion across Heterogeneous Ontology Alignment Techniques
A two-stage ensemble reconciles outputs from string, graph-embedding, and RAG-based ontology aligners. Composition matters: mixed paradigms favor precision, while LLM-only groups more often favor F1.
OntoAligner-Ensemble applies **voting-based fusion** and then post-fusion selection to candidate correspondences from string, KGE, and RAG-based aligners. It was evaluated on **eight tasks across five OAEI tracks**.
For pipelines that reconcile structured data, treat ensemble composition as a tunable policy. Cross-paradigm groups generally favored precision, while **homogeneous LLM ensembles** more often delivered higher overall F1.
OntoAligner-Ensemble applies **voting-based fusion** and then post-fusion selection to candidate correspondences from string, KGE, and RAG-based aligners. It was evaluated on **eight tasks across five OAEI tracks**. For pipelines that reconcile structured data, treat ensemble composition as a tunable policy. Cross-paradigm groups generally favored precision, while **homogeneous LLM ensembles** more often delivered higher overall F1. The material provides no per-task scores, costs, or latency results. The findings are specific to ontology alignment, so their value for general coding-agent orchestration remains untested.
This narrows multi-agent ensemble claims to a specific fusion policy and domain: composition and post-fusion selection change the precision–F1 tradeoff, and heterogeneous ensembles are not automatically superior. Against the prior candidates, it supports typed, tunable aggregation rather than unconstrained agent debate, while leaving general orchestration value, coherence, cost, and latency unresolved.