Query-Based Entity Comparison in Knowledge Graphs Revisited

Large-scale knowledge graphs are increasingly being used in applications, and there is a growing need for tools that can effectively support users in analysis and exploration tasks. One such important task is entity comparison-to describe in an informative way the similarities between two given entities as described in a knowledge graph. In our previous work the result of entity comparison is modelled as a similarity query-that is, a SPARQL query having the input entities as part of the answer over the input graph; for instance, one can describe the similarity between two companies such as Telenor and Vodafone in the YAGO graph as a query asking for all telecom companies based in Europe. In this paper, we extend the results of our prior work in different ways. First, we expand the language of similarity queries to consider a richer fragment of SPARQL allowing for numeric filter expressions; this enables us to express that Telenor and Vodafone are also similar in that they both have at least 30,000 employees. We then propose algorithms for computing similarity queries satisfying certain additional desirable properties, such as being as specific as possible. Such algorithms are, however, impractical; hence, we also propose and implement a scalable algorithm that is guaranteed to compute a similarity query, but not necessarily a most specific one.