Efficient cross-validation in ILP

Abstract

Cross-validation is a technique used in many different machine learning approaches. Straightforward implementation of this technique has the disadvantage of causing computational overhead. However, it has been shown that this overhead often consists of redundant computations, which can be avoided by performing all folds of the crossvalidation in parallel. In this paper we study to what extent such a parallel algorithm is also useful in ILP. We discuss two issues: a) the existence of dependencies between parts of a query that limit the obtainable efficiency improvements and b) the combination of parallel cross-validation with query-packs. Tentative solutions are proposed and evaluated experimentally.