A Pattern Synthesis Technique To Reduce The Curse Of Dimensionality Effect

Abstract

The number of samples needed to estimate a function or model grows enormously as the dimensionality of the feature space increases. This is called the curse of dimensionality effect. As a result with limited or small training set the estimated model can become severely biased for high dimensional data. There exists two broad ways to tackle this problem, viz., feature selection and bootstrapping. Feature selection tries to reduce the dimensionality of the data whereas bootstrapping are basically resampling techniques. We propose a novel solution for this problem based on a pattern synthesis technique. An instance based pattern synthesis technique called overlap based pattern synthesis which can generate exponential number of synthetic patterns along with a compact representation of the entire synthetic set are presented. These techniques are applied with a nearest neighbor classifier (NNC) called OLP-NNC which directly works with the compact representations and experimentally demonstrated that not only its classification accuracy is increased but its computational requirements are reduced when compared with conventional NNC and other relevant methods.