Utilization of dimensionality reduction in stacked generalization architecture

Stacked generalization (SG) is a hierarchical architecture, which combines classifiers in order to boost the performance of overall system by integrating the individual classifiers. Two-layered version of SG has been utilized in many image analysis researches and shown to be an effective tool for enhancing the performance measure of the individual classifiers in a number of various cases. However, due to its layered architecture and integration approach, it inherits some problems such as curse of dimensionality, i.e. necessity for more samples, and higher computational time for classification. In this work, the effect of two different dimensionality reduction methods, namely linear and non-linear, between layers of SG are analyzed to attack the curse of dimensionality problem. In the experimental part, the comparisons of these approaches with respect to each other and SG-without-dimensionality-reduction are presented. According to test results, it is concluded that dimensionality reduction has a positive effect on the performance of SG, i.e. both linear and nonlinear dimensionality reduction approaches performs better than SG architecture, and nonlinear dimensionality reduction achieves better classification accuracy than linear dimensionality reduction.