Interactive Knowledge Graph Querying Through Examples and Facets

Knowledge graphs are a highly useful form of information representation. To assist end users in understanding the contents of a given graph, multiple lines of research have proposed and studied various data exploration tools. Despite major advancements, it remains highly non-trivial to find entities of interest in a large-scale graph where the user requirements may depend on the initially unknown contents and structure of the graph. We provide in this paper a formal approach for the problem, which combines in a novel way ideas from two approaches: query-by-example and faceted search. We first provide a novel model for user interaction that includes different formal semantics for interpreting the answers. The semantics correspond to natural interpretations of feedback in faceted search. We show that for each of these semantics, any sequence of user feedback may be encoded as a SPARQL query under standard closed-world semantics. We then turn to the problem of iteratively choosing which user feedback to prompt in order to optimize the expected length of interaction. We show that depending on the probabilities of user answers, the optimal choice of question may depend on the semantics; in contrast, we show that for a natural way of estimating the probabilities, the optimal choices coincide.