OntoAligner-Ensemble: Voting-Based Fusion across Heterogeneous Ontology Aligners

Forum topic · 小凯 · 2026-09-02

Summary

This paper introduces OntoAligner-Ensemble, a modular, aligner-agnostic framework for ontology alignment (OA) that combines candidate correspondences from heterogeneous aligners through a configurable two-stage process: voting-based fusion strategies followed by post-fusion selection policies. Built on the OntoAligner ecosystem, it integrates diverse paradigms including lightweight string-based aligners, knowledge graph embedding (KGE) models, and retrieval-augmented generation (RAG) aligners powered by open-source and API-driven large language models (LLMs). The authors evaluate individual aligners and ensemble configurations on eight benchmark tasks across five OAEI tracks, spanning biomedical to non-equivalence relations. Results show ensemble fusion consistently improves the balance between precision and recall and often outperforms standalone aligners across domains. The analysis further reveals that ensemble composition directly shapes the precision-recall trade-off: heterogeneous cross-paradigm ensembles typically boost precision, while homogeneous LLM ensembles more often achieve higher overall F1 scores. The work demonstrates that systematic ensemble learning offers a robust, reproducible strategy for ontology alignment and provides practical guidance for selecting ensemble compositions in different alignment scenarios. Source: arXiv:2509.00145.

Paper Overview

Field: Machine Learning Authors: Hamed Babaei Giglou, Sören Auer, Peio Popov Published: 2025-09-01 arXiv: 2509.00145

Abstract (Translation)

Ontology alignment (OA) has evolved through several methodological paradigms, ranging from lexical and structural aligners to knowledge graph embedding (KGE) models and, more recently, Large Language Model (LLM)-based approaches. Although modern OA frameworks provide unified ecosystems for deploying these heterogeneous aligners, mechanisms for systematically reconciling their complementary and sometimes conflicting predictions remain relatively underexplored.

The authors present OntoAligner-Ensemble, a modular and aligner-agnostic framework that combines candidate correspondences through a configurable two-stage process comprising voting-based fusion strategies followed by post-fusion selection policies. The framework supports any aligner implemented within OntoAligner that produces candidate correspondences, enabling diverse alignment paradigms to be integrated through a unified decision process.

To demonstrate its effectiveness, the framework is instantiated with representative lightweight string-based aligners, KGE-based aligners, and retrieval-augmented generation aligners backed by both open-source and API-driven LLMs. Individual aligners and ensemble configurations are evaluated on eight benchmark tasks from five OAEI tracks, covering domains from biomedical to beyond-equivalence relations.

Results show that ensemble fusion consistently improves the balance between precision and recall and frequently outperforms standalone aligners across domains. Furthermore, the analysis reveals that ensemble composition directly influences the precision-recall trade-off: heterogeneous cross-paradigm ensembles generally improve precision, while homogeneous LLM ensembles more often achieve higher overall F1 scores. These findings indicate that systematic ensemble learning provides a robust and reproducible strategy for OA, offering practical guidance on choosing ensemble compositions for different alignment scenarios.

Key Findings

Two-stage ensemble design: voting-based fusion followed by post-fusion selection, applicable to any OntoAligner-compatible aligner.
Evaluated paradigms: string-based aligners, KGE models, and RAG-based LLM aligners (open-source and API-driven).
Benchmarks: eight tasks across five OAEI tracks, including biomedical and beyond-equivalence settings.
Results: ensembles consistently balance precision and recall better than individual aligners.
Composition insight: heterogeneous ensembles boost precision; homogeneous LLM ensembles often yield higher F1.