Learning time series models with inductive logic programming

Abstract

: This paper reports on a set of proposals that make Inductive Logic Programming (ILP) systems adequate for inducing time series models. The proposals include an improvement in the ILP search process by the introduction of a statistical model validation step. We propose the definition of an adequate cost function based on the information criteria. The definition of the model evaluation step consists in an intuitive statistics that limits the minimum accepted performance of an induced hypothesis. The ILP system we used was provided with a library of background knowledge predicates adequate for time series problems. The proposals described in this paper can be applied to any agnostic learning problem. Preliminary experiments have shown that all these modifications make an ILP system adequate to induce time series models and increase the capability of model choice automation.