Class-specific feature generation for 1NN through genetic programming

This paper introduces a genetic program for class-specific feature extraction for 1NN. Under the proposed method a new feature space is generated for each class in the problem under analysis. Where feature spaces are build by merging the initial features with a genetic program that aims at maximizing classification accuracy of a 1NN classifier. We compare the performance of our method to both, classical-standard techniques (e.g., PCA, LDA) and to solutions based on evolutionary algorithms. Experimental results reveal our method outperforms alternative solutions in a wide variety of data sets.