SchemaML: Semantic Schema Matching for Enhanced Data Interoperability
Abstract
Schema matching plays a crucial role in data integration by aligning attributes from heterogeneous data sources. However, traditional approaches often fail to capture semantic relationships between attributes with different naming conventions. To address this limitation, we propose SchemaML, a supervised machine learning approach that combines semantic word embeddings with a Random Forest classifier. The proposed method follows a structured pipeline including column extraction, embedding generation using pre-trained models, pairwise feature construction, and supervised classification. Experiments are conducted on datasets derived from schema.org, covering multiple schema sizes (10, 20, and 50 columns) to evaluate scalability and robustness. SchemaML achieves an accuracy of up to 94%, with precision and recall exceeding 93%, outperforming traditional rule-based and machine learning baselines under identical experimental conditions. Additional evaluations demonstrate robustness to noisy data, including typographical errors and abbreviations, with performance remaining stable under 10% noise injection. These results highlight the effectiveness of combining semantic representations and ensemble learning for scalable and automated schema matching in real-world data integration scenarios.