Biasfree Hypothesis Evaluation in Multirelational Domains.pdfVIP

Bias-free Hypothesis Evaluation in Multirelational Domains ∗ ¨ Christine Korner Stefan Wrobel Fraunhofer Institut AIS, Germany Fraunhofer Institut AIS and christine.koerner@ais.fraunhofer.de Dept. of Computer Science III University of Bonn, Germany stefan.wrobel@ais.fraunhofer.de Abstract 2002] eliminates this effect but assumes that information shared in the training set does not reoccur with future in- In propositional domains, using a separate test stances of the domain. We introduce generalized subgraph set via random sampling or cross validation is sampling which prevents a bias due to an increased or re- generally considered to be an unbiased estimator duced proportion of linked objects in the test set and which of true error. In multirelational domains, previ- includes subgraph sampling as a special case. ous work has already noted that linkage of ob- The paper is organized as follows. Section 2 gives an jects may cause these procedures to be biased, overview of commonly used methods for error estimation and has proposed corrected sampling procedures. on propositional data and shows by a simple experiment However, as we show in this paper, the exist- the different behavior of true erro