Schema matching prediction with applications to data source discovery and dynamic ensembling

摘要: Web-scale data integration involves fully automated efforts which lack knowledge of the exact match between descriptions. In this paper, we introduce schema matching prediction, an assessment mechanism to support matchers in absence match. Given attribute pair-wise similarity measures, a predictor predicts success matcher identifying correct correspondences. We present comprehensive framework predictors can be defined, designed, and evaluated. formally define evaluation prediction using spaces discuss set four desirable properties predictors, namely correlation, robustness, tunability, generalization. method for constructing supporting generalization, models as means tuning toward various quality measures. empirical correlation robustness provide concrete measures their evaluation. illustrate usefulness by presenting three use cases: propose ranking relevance deep Web sources with respect given user needs. show how assist design systems. Finally, dynamic weight setting ensemble, thus improving upon current state-of-the-art methods. An extensive shows these cases demonstrates increasing performance matching.