Abducing Rules with Predicate Invention

This paper addresses discovery of unknown relations from incomplete network data using abduction. Given a network information such as causal relations and metabolic pathways, we want to infer missing links and nodes in the network to account for observations. This is implemented in SOLAR, an automated deduction system for consequence finding, using first-order representation of algebraic relations and full-clausal ground formulas of network information. Abduction by SOLAR is powerful enough to infer unknown rules and to realize predicate invention by inferring unknown causes. In particular, we point out the importance of existentially quantified formulas to express hypotheses including new variables representing missing nodes to be introduced.