Evolving linear neural networks for features space dimensionality reduction

Principal Components Analysis (PCA) is one of the most wide-spread methods for dimensionality reduction, which is being applied in many research and problem domains. So far a lot of approaches to compute data matrix eigenvectors, which correspond to the Principal Components, were proposed, among which numerical methods and Hebbian-based learning for neural networks, including Generalized Hebbian Algorithm. In this paper a novel way for computing eigenvectors using evolving linear neural networks is introduced, which is not relying upon correlation between nodes, but uses special fitness function instead. Early removal of the low-informative linear subspaces is applied, which reduces computational complexity of the method, and besides eigenvectors coordinates are computed approximately to improve convergence and speed. The latter gave rise to the approach´s name: pseudo-PCA. Experimental results show that not looking at inexact eigenvectors the approach allows effective reduction of the features space dimensionality with acceptable classification accuracy compared to some “classical” and modern approaches to solve classification problems.