REGNUM: Generating Logical Rules with Numerical Predicates in Knowledge Graphs

Mining logical rules from a knowledge graph (KG) can reveal useful patterns for predicting facts, curating the KG, and identifying trends. However, many rule mining systems face challenges when working with numerical data because numerical predicates can take a large number of values, leading to a huge search space. In this work, we present REGNUM, a system that addresses this issue by generating rules with numerical constraints. REGNUM extends the body of rules mined from a KG by using supervised discretization of numerical values with decision trees to increase the confidence of the rules without sacrificing significance. Our experimental results show that the numerical rules have a higher overall quality than the parent rules and are effective at making better predictions.