A performance comparative analysis between rule-induction algorithms and clustering-based constructive rule-induction algorithms. Application to rheumatoid arthritis

Abstract

We present a performance comparative analysis between traditional rule-induction algorithms and clustering-based constructive rule-induction algorithms. The main idea behind these methods is to find dependency relations among primitive variables and use them to generate new features. These dependencies, corresponding to regions in the space, can be represented as clusters of examples. Unsupervised clustering methods are proposed for searching for these dependencies. As a benchmark, a database of rheumatoid arthritis (RA) patients has been used. A set of clinical prediction rules for prognosis in RA was obtained by applying the most successful methods, selected according to the study outcomes. We suggest that it is possible to relate predictive features and long-term outcomes in RA. © Springer-Verlag Berlin Heidelberg 2004.