Filtering Recommender System for Semantic Model Refinement

Alexander Paulus, Andreas Burgdorf, André Pomp, Tobias Meisen

Abstract: In order to prepare data for a semantic data integration (e.g., into knowledge graphs), a semantic mapping in the form of a semantic model between data and the corresponding ontology is necessary. Manual mapping, although a time-consuming process, is often required as the models created by automated approaches still need improvements. For supporting this refinement process, we propose a semi-automatic modeling process that includes the use of a recommender system. This recommender system supports the modeler in the context of a manual refinement process by suggesting possible adjustments consisting of new nodes and corresponding relations to the model. In our approach, we use random forest classifiers and graph neural networks to generate recommendations for model modifications following a collaborative filtering idea and using information from existing semantic models. We evaluated our approach on different public datasets. We achieve an MRR of 0.69 and Hits@3 of 0.86 for our predictions showing that the approach is able to provide suitable recommendations to support the modeler based on the characteristics of the training data.