Decomposition of high dimensional pattern spaces for hierarchical classification

In this paper we present a novel approach to decomposing high dimensional s paces using a multiobjective genetic algorithm for identifying (near-)optim al subspaces for hierarchical classification. This strategy of pre-processi ng the data and explicitly optimising the partitions for subsequent mapping onto a hierarchical classifier is found to both reduce the learning comple xity and the classification time with no degradation in overall classificat ion error rate. Results of partitioning pattern spaces are presented and co mpared with various algorithms.