Do you catch my drift? On the usage of embedding methods to measure concept shift in knowledge graphs

Automatically detecting and measuring differences between evolving Knowledge Graphs (KGs) has been a topic of investigation for years. With the rising popularity of embedding methods, we investigate the possibility of using embeddings to detect Concept Shift in evolving KGs. Specifically, we go deeper into the usage of nearest neighbour set comparison as the basis for a similarity measure, and show why this approach is conceptually problematic. As an alternative, we explore the possibility of using clustering methods. This paper serves to (i) inform the community about the challenges that arise when using KG embeddings for the comparison of different versions of a KG specifically, (ii) investigate how this is supported by theories on knowledge representation and semantic representation in NLP and (iii) take the first steps into the direction of valuable representation of semantics within KGs for comparison.