Evolving linear discriminant in a continuously growing dimensional space for incremental attribute learning

Abstract

Feature Ordering is a unique preprocessing step in Incremental Attribute Learning (IAL), where features are gradually trained one after another. In previous studies, feature ordering derived based upon each individual feature's contribution is time-consuming. This study attempts to develop an efficient feature ordering algorithm by some evolutionary approaches. The feature ordering algorithm presented in this paper is based on a criterion of maximum mean of feature discriminability. Experimental results derived by ITID, a neural IAL algorithm, show that such a feature ordering algorithm has a higher probability to obtain the lowest classification error rate with datasets from UCI Machine Learning Repository. © IFIP International Federation for Information Processing 2012.