Feature Subset Selection and Inductive Logic Programming

Abstract

Applicability of ILP to real-world problems is constrained by the high dimensionality of ILP tasks. This paper proposes to reduce the dimensionality of the ILP example space by bringing feature subset selection to the realm of ILP. Seen as a black box, the method reduces ILP examples, in the form of nonrecursive Datalog clauses, by removing literals judged irrelevant for the ILP task at hand. The approach exploits existing AVL feature selection (FS) algorithms by changing the representation from the FOL clausal format to a fixed-length AV format, performing FS on the AV approximation of the FOL task, and mapping the results back to the FOL representation. Issues of noise introduced in the data during representation shift are presented, and constraints on the AVL FS to be used in ILP FS are discussed.