OntoAligner-Ensemble: Voting-Based Fusion across He...

OntoAligner-Ensemble: Voting-Based Fusion across He... | AI Research

Key Takeaways

OntoAligner-Ensemble is a modular framework designed to reconcile the predictions of diverse ontology alignment (OA) techniques.
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, we instantiate the framework using representative lightweight string-aligner, KGE-based, and Retrieval-Augmented Generation aligners powered by both open-weight and API-based LLMs.
We evaluate individual aligners and ensemble configurations across eight benchmark tasks from five OAEI tracks spanning biomedical to beyond-equivalence.
The results show that ensemble fusion consistently improves the balance between precision and recall and frequently outperforms standalone aligners across diverse domains.

Paper AbstractExpand

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. We 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, we instantiate the framework using representative lightweight string-aligner, KGE-based, and Retrieval-Augmented Generation aligners powered by both open-weight and API-based LLMs. We evaluate individual aligners and ensemble configurations across eight benchmark tasks from five OAEI tracks spanning biomedical to beyond-equivalence. The results show that ensemble fusion consistently improves the balance between precision and recall and frequently outperforms standalone aligners across diverse domains. Furthermore, our analysis reveals that ensemble composition directly affects the precision-recall trade-off: heterogeneous cross-paradigm ensembles generally improve precision, whereas homogeneous LLM ensembles more often achieve higher overall F1-scores. These findings demonstrate that systematic ensemble learning offers a robust and reproducible strategy for OA while providing practical guidance for selecting ensemble compositions under different alignment scenarios.

OntoAligner-Ensemble is a modular framework designed to reconcile the predictions of diverse ontology alignment (OA) techniques. By using a voting-based fusion process, it allows researchers to combine different alignment paradigms—such as lexical, structural, knowledge graph embedding (KGE), and Large Language Model (LLM) approaches—into a single, unified decision process. The framework aims to mitigate the individual weaknesses of these models while leveraging their complementary strengths across various domains.

Addressing Alignment Variability

Ontology alignment involves identifying semantic correspondences between concepts in different ontologies. Historically, this field has evolved through distinct paradigms, including rule-based, similarity-based, and machine learning-based methods. However, no single paradigm consistently performs best across all domains. Lexical matchers may struggle with semantic drift, KGE-based models can be sensitive to graph sparsity, and LLM-based aligners vary in performance depending on the model and provider. OntoAligner-Ensemble addresses this by providing a systematic way to integrate these heterogeneous aligners, ensuring that their conflicting or complementary outputs are reconciled through a configurable, aligner-agnostic pipeline.

How the Framework Works

The framework operates through a two-stage process:

Fusion Stage: Multiple aligners generate candidate correspondences. These are aggregated using voting strategies such as Weighted Voting, Reciprocal Rank Fusion, Condorcet, or Borda Count. Each aligner can be assigned a weight to reflect its reliability, and the system ensures that duplicate candidates are resolved by retaining the highest-scoring occurrence. 2. Selection Stage: A post-fusion selection policy converts the ranked candidates into a final alignment. Available policies include Top-1, Top-k, threshold-based filtering, and greedy bijective selection for one-to-one mapping requirements.
This modular design allows the framework to remain independent of the underlying algorithms, supporting any aligner that can produce candidate correspondences within the OntoAligner ecosystem.

Performance and Ensemble Composition

The authors evaluated the framework across eight benchmark tasks from five OAEI tracks, including biomedical, material science, biodiversity, and beyond-equivalence settings. The results indicate that ensemble fusion consistently improves the balance between precision and recall compared to standalone aligners.
The study also identified a clear trade-off based on ensemble composition:

Heterogeneous Ensembles: Combining different alignment paradigms (e.g., mixing lexical, KGE, and LLM-based aligners) generally improves precision.
Homogeneous Ensembles: Using a group of similar models, such as multiple LLM-based aligners, more often results in higher overall F1-scores.

Considerations for Implementation

While ensemble learning offers a robust strategy for improving alignment quality, the authors note that it introduces additional computational overhead. Combining multiple aligners increases execution time and resource requirements compared to running a single model. Consequently, this approach is most suitable for scenarios where accuracy and the ability to handle heterogeneous ontologies are prioritized over real-time response efficiency.