Evolving Kernel PCA Pipelines with Evolution Strategies

Abstract

This paper introduces an evolutionary tuning approach for a pipeline of preprocessing methods and kernel principal component analysis (PCA) employing evolution strategies (ES). A simple (1+1)-ES adapts the imputation method, various preprocessing steps like normalization and standardization, and optimizes the parameters of kernel PCA. A small experimental study on a benchmark data set with missing values demonstrates that the evolutionary kernel PCA pipeline can be tuned with relatively few optimization steps, which makes evolutionary tuning applicable to scenarios with very large data sets.