Learning Knowledge Graph for Target-driven Schema Matching

Though research on schema matching goes back many years, this remains a largely unsolved problem on the ground. The knowledge of concepts, terms and their correspondences useful for matching is implicit in historical matching information. Based on this insight, we propose a probabilistic framework for schema matching based on learning. It updates a knowledge graph of concepts, terms, their relationships and associated probabilities from historical `training' matches, and then uses this knowledge for predicting matches for a `test' source schema. We present a solution based on stochastic EM, where the latent concepts and relationships in historical matches are inferred by sampling during the expectation step, and parameters associated with the knowledge graph are estimated using these inferred variables. Using extensive experiments we show that by making effective use of historical matches, our approach significantly outperforms state-of-the-art matching techniques for a real-world Insurance dataset as well as a benchmark dataset.