BioGITOM: Matching Biomedical Ontologies with Graph Isomorphism Transformer

BioGITOM: Matching Biomedical Ontologies with Graph Isomorphism Transformer

연구 분야: Software Development

학회: The VLDB Journal

Ontology Matching (OM) plays a central role in ensuring interoperability across heterogeneous biomedical ontologies. Existing approaches are broadly classified into (i) traditional methods that rely on external lexicons and predefined rules, and (ii) learning methods that leverage Deep Learning (DL) and Graph Neural Networks (GNNs) to generate expressive concept representations. In particular, learning OM methods often make use of Graph Convolutional Networks. Motivated by the success of Graph Isomorphism Networks and the versatility of Graph Transformers across various applications, in this paper, we propose a hybrid GNN model named Graph Isomorphism Transformer (GIT). This study explores the benefits of applying the GIT model to OM, emphasizing its potential to enhance accuracy, and improve the scalability of learning-based systems. We introduce BIOGITOM, a new OM approach comprising five core modules: (1) the Preprocessing is applied to refine raw data and extract pertinent features; (2) the Concept Features Encoder, which generates semantic encodings; (3) the GIT model, tailored to enhance concepts embeddings with structural features; (4) the Gating Aggregator, employed to derive final concepts’ embeddings by integrating both semantic and structural feature encodings; and (5) the Mappings Selector, designed to identify mappings between concepts. Comprehensive experiments conducted on the Bio-ML track of the Ontology Alignment Evaluation Initiative ( ) showcase the effectiveness of BIOGITOM. The results highlight the superior performance of BIOGITOM compared to state-of-the-art traditional and learning-based methods.