Semantic Concept Recommendation for Continuously Evolving Knowledge Graphs

André Pomp, Vadim Kraus, Lucian Poth, Tobias Meisen

Abstract: With the digitalization of many industrial processes and the increasing interconnection of devices, the number of data sources and associated data sets is constantly increasing. Due to the heterogeneity of these large amounts of data sources, finding, accessing and understanding them is a major challenge for data consumers who want to work with the data. In order to make these data sources searchable and understandable, the paradigms of Ontology-Based Data Access (OBDA) or Ontology-Based Data Integration (OBDI) are used today. An important part of these paradigms is the creation of a mapping, such as a semantic model, between a previously defined ontology and the existing data sources. Although there are already many approaches that automate the creation of this mapping by using data-driven or data-structure-driven approaches, none of them focuses on the fact that the underlying ontology evolves over time. However, this is essential in today’s age of large amounts of data and ever-growing number of data sources. In this paper, we propose an approach that allows the recommendation of semantic concepts for data attributes based on a constantly evolving knowledge graph. The approach allows the knowledge graph to learn data-driven representations for any concept that is available in the knowledge graph and that is already mapped to at least one data attribute. Instead of supporting a single method for recommending semantic concepts, we design the approach to be able to learn multiple data representatives per semantic concept, with each representative being trained on a different method, such as machine learning classifiers, rules, or statistical methods. In this way, for example, we are able to distinguish between different data types and data distributions. In order to evaluate our approach, we have trained it on several different publicly available data sets. In comparison to existing approaches, our evaluation shows that the accuracy of the recommendation improves through the use of our flexible and dedicated classification approach.