Partitioning the Input Domain for Classification

Abstract

We explore an approach to use simple classification models to solve complex problems by partitioning the input domain into smaller regions that are more amenable to the classifier. For this purpose weinvestigate two variants of partitioning based on energy, as measured by the variance. We argue that restricting the energy of the input domain limits the complexity of the problem. Therefore, our method directly controls the energy in each partition. The partitioning methods and several classifiers are evaluated on a road detection application. Our results indicate that partitioning improves the performance of a linear Support Vector Machine and a classifier which considers the average label in each partition, to match the performance of a more sophisticated Neural Network classifier.