Abstract:

Aiming at the problem of high labeling cost and imbalanced data exists in the machine learning based ontology matching, the model ontology matching is taken as an co-training problem on two views, which respectively extracts feature sets from the ontology schema level and the data level. Through pre-matching ontologies’ concept pairs, the imbalanced degree of samples is reduced. After analyzing the limitations of the traditional co-training method, in combination with the active learning theory, an improved sample value considered co-training method is proposed, it can choose more valuable unlabeled samples to update training sets in each iteration. Experimental results show that the proposed method is more efficient and achieves better competitive ontology matching results.