Ranking Indicator Discovery from Very Large Knowledge Graphs

Download PDF

Abstract

Ranking indicators are essential tools for comparing the importance of various entities such as cities or scientists. While extensively used in fields like econometrics and scientometrics, many other domains lack systematic approaches for developing these indicators. In this paper, we introduce a novel method for automatically discovering ranking indicators from very large knowledge graphs. To this end, we formalize the notion of counting graph pattern (CG) as a special SPARQL query, and the concept of ideal ranking indicator as a CG whose result induces a strict total order on a set of entities. To assess the interestingness of ranking indicators, we employ the proportion of covered entities along with an inequality measure, namely the Gini coefficient. We further present Algorithm Ranking Indicator Pattern Miner ( RIPM ) , to efficiently identify interesting ranking indicators for a given field, thanks to pruning techniques for handling the very large search space. Our experimental study shows the effectiveness of our optimizations. It also validates that RIPM extracts transparent, diverse, and understandable indicators through a user survey and a comparison with two baselines. This work has significant implications for fields lacking dedicated communities working on ranking tasks, providing a robust tool to automatically produce ranking indicators, and the associated rankings.