Enhancing Knowledge Graphs with Data Representatives

Abstract

Due to the digitalization of many processes in companies and the increasing networking of devices, there is an ever-increasing amount of data sources and corresponding data sets. To make these data sets accessible, searchable and understandable, recent approaches focus on the creation of semantic models by domain experts, which enable the annotation of the available data attributes with meaningful semantic concepts from knowledge graphs. For simplifying the annotation process, recommendation engines based on the data attribute labels can support this process. However, as soon as the labels are incomprehensible, cryptic or ambiguous, the domain expert will not receive any support. In this paper, we propose a semantic concept recommendation for data attributes based on the data values rather than on the label. Therefore, we extend knowledge graphs to learn different dedicated data representations by including data instances. Using different approaches, such as machine learning, rules or statistical methods, enables us to recommend semantic concepts based on the content of data points rather than on the labels. Our evaluation with public available data sets shows that the accuracy improves when using our flexible and dedicated classification approach. Further, we present shortcomings and extension points that we received from the analysis of our evaluation.