An Approach to Predicate Invention Based on Statistical Relational Model

Abstract

Predicate Invention is the branch of symbolic Machine Learning aimed at discovering new emerging concepts in the available knowledge. The outcome of this task may have important consequences on the efficiency and effectiveness of many kinds of exploitation of the available knowledge. Two fundamental problems in Predicate Invention are how to handle the combinatorial explosion of candidate concepts to be invented, and how to select only those that are really relevant. Due to the huge number of candidates, there is a need for automatic techniques to assign a degree of relevance to the various candidates and select the best ones. Purely logical approaches may be too rigid for this purpose, while statistical solutions may provide the required flexibility.

This paper proposes a new Statistical Relational Learning approach to Predicate Invention. The candidate predicates are identified in a logic theory, rather than in the background knowledge, and are used to restructure the theory itself. Specifically, the proposed approach exploits the Markov Logic Networks framework to assess the relevance of candidate predicate definitions. It was implemented and tested on a traditional problem in Inductive Logic Programming, yielding interesting results.