TermPicker : Recommending Vocabulary Terms for Reuse When Modeling Linked Open Data
Reusing terms from Resource Description Framework (RDF) vocabularies when modeling data as Linked Open Data (LOD) is difficult and without additional guidance far from trivial. This work proposes and evaluates TermPicker: a novel approach alleviating this situation by recommending vocabulary terms based on the information how other data providers modeled their data as LOD. TermPicker gathers such information and represents it via so- called schema-level patterns (SLPs), which are used to calculate a ranked list of RDF vocabulary term recommendations. The ranking of the recommendations is based either on the machine learning approach "Learning To Rank" (L2R) or on the data mining approach "Association Rule" mining (AR). TermPicker is evaluated in a two-fold way. First, an automated cross-validation evaluates TermPicker’s prediction based on the Mean Average Precision (MAP) as well as the Mean Reciprocal Rank at the first five positions (MRR@5). Second, a user study examines which of the recommendation methods (L2R vs. AR) aids real users more to reuse RDF vocabulary terms in a practical setting. The participants, i.e., TermPicker’s potential users, are asked to reuse vocabulary terms while modeling three data sets as LOD, but they receive either L2R-based recommendations, AR-based recommendation, or no recommendations. The results of the cross-validation show that using SLPs, TermPicker achieves 35% higher MAP and MRR@5 values compared to using solely the features based on the typical reuse strategies. Both the L2R-based and the AR-based recommendation methods were able to calculate lists of recommendations with MAP = 0.75 and MRR@5 = 0.80. However, the results of the user study show that the majority of the participants favor the AR-based recommendations. The outcome of this work demonstrates that TermPicker alleviates the situation of searching for classes and properties used by other data providers on the LOD cloud for representing similar data.