Relational Concepts and the Fourier Transform: An Empirical Study

摘要

Lack of domain knowledge may impose primitive data representation. Then, compelx (non-linear) relationships among attributes complicate learning, especially for typical learning methods. These methods fail because their bias does not match the compelx relational structures relevenat to the domain. However, more recent approaches to learning have implemented biases that allow learning of structured, albeit complex, concepts. One of such approaches, based o nthe Fourier transform of Boolean functions, is studied and compared empirically to others, based on constructing new features or extracting relations from propositional training data. Controlled experiments help to characterized the kinds of concept that allow each approach to outperform the others. This characterization, which implicates parameters of Fourier complexity, other measures of concept difficulty, and the relational strucure of the target concepts, is also discussed with respect to difficult real-world domains.