CoEPinKB: A Framework to Understand the Connectivity of Entity Pairs in Knowledge Bases

Abstract

A knowledge base, expressed using the Resource Description Framework (RDF), can be viewed as a graph whose nodes represent entities and whose edges denote relationships. The entity relatedness problem refers to the problem of discovering and understanding how two entities are related, directly or indirectly, that is, how they are connected by paths in a knowledge base. Strategies designed to solve the entity relatedness problem typically adopt an entity similarity measure to reduce the path search space and a path ranking measure to order and filter the list of paths returned. This paper presents a framework, called CoEPinKB, that supports the empirical evaluation of such strategies. The proposed framework allows combining entity similarity and path ranking measures to generate different path search strategies. The main goals of this paper are to describe the framework and present a performance evaluation of nine different path search strategies.

A Framework to Compute Entity Relatedness in Large RDF Knowledge Bases

“…Although the experiments used a standalone implementation, the framework can be retargeted to a fully distributed environment. By contrast, the CoEPinKB framework [Jiménez et al 2021] implements path search strategies using a multi-thread approach. Table IV compares DCoEPinKB with related systems.…”

“…The Spark Query Executor component interacts with DataFrames that represent views of RDF datasets stored in the distributed file system as Parquet files. Based on previous experience Jiménez et al [2021], we implemented translations of the required SPARQL queries to Spark SQL queries. Spark SQL is one of the most popular modules of Spark, targeted for processing structured data, using the Datasets and DataFrames data abstractions, and provides support for reading and writing Parquet files.…”

“…Figure 3a shows the average execution times of the J&E strategy, which achieved the best performance for finding relevant relationship paths in Herrera [2017], and the best average execution time in Jiménez et al [2021]). For each pair of entities in each dataset, we searched 6 times for the top-50 paths between them, excluded the first run time to avoid the warm-up bias, and calculated the average time of the last five executions.…”

“…Following Fang et al [2011], Herrera [2017] and Jiménez et al [2021], this article then investigates the entity relatedness problem for large RDF knowledge bases by exploring path search strategies, which have two major steps. The first step uses the Backward Search Heuristic (BSH) [Le et al 2014], which is a breadth-first search strategy that expands the paths starting from each input entity, until a candidate relationship path is generated.…”

“…The article describes a framework, called DCoEPinKB, which allows experimenting with various path search strategies over large RDF graphs, using different entity similarity measures, expansion limits, and path ranking measures. First introduced in [Guillot Jiménez et al 2021] and unlike the approaches described in Herrera [2017] and Jiménez et al [2021], DCoEPinKB is built on top of Apache Spark [Zaharia et al 2010]. The experiments use a proof-of-concept standalone setup, leaving to future work testing the framework in a fully distributed environment.…”