Connecting the Dots: Explaining Relationships Between Unconnected Entities in a Knowledge Graph Nitish Aggarwal ∗ , Sumit Bhatia ◦ , and Vinith Misra ◦ ∗ Insight Centre for Data Analytics, National University of Ireland, Galway, Ireland ◦ IBM Research, Almaden, USA nitish.aggarwal@insight-center.org,{sumit.bhatia,vmisra}@us.ibm.com Abstract. We discuss the problem of explaining relationships between two un- connected entities in a knowledge graph. We frame it as a path ranking problem and propose a path ranking mechanism that utilizes features such as specificity, connectivity, and path cohesiveness. We also report results of a preliminary user evaluation and discuss a few example results. 1 Introduction The advent of semantic knowledge bases like DBpedia, Freebase, etc. has led to the development of smart search systems that produce rich and enhanced results by pro- viding additional related information about the entities/concepts being queried by the users. Further, increasing efforts are being made to build knowledge discovery systems that help users to navigate/explore the semantic graph and discover hitherto unknown, yet extremely useful information. For example, Nagarajan et al. [6] describe a discovery system that uses a semantic network built out of medical literature and helps researchers in discovering previously unknown protein-protein interactions. Likewise, web search engines like Google, Bing, etc. also incorporate data from their knowledge graphs to provide a list of entities that are related to the user search query [3] and users often navigate through these recommended entities to discover new non-trivial information about their search topics. Despite providing a greatly simplified knowledge discovery process by recommend- ing related entities of interest, such systems often fail to provide explanations for such recommendations to users, especially for less popular entities and entities that are not directly connected to the input entity. For example, for an entity query “Abu Bakr al-Baghdadi”(leader of the terrorist organization Islamic State of Iraq and the Levant (ISIL)), Google recommends entities such as “Musab al-Zarqawi”, “Qasem Soleimani”, etc., but fails to provide any explanation about how these entities are related to the in- put entity. Previous research efforts [5,7] have tried to explain the relatedness between entities by deriving important paths between entities in the knowledge graph. However, these methods generally focus either on popular entities in the graph or rely on query log data from the search engines that may not be available always, especially in enterprise domains.