Counter-clustering for Training Pattern Selection
Tatiana Tambouratzis · The Computer Journal · 2000
Training pattern selection consists of selecting a subset of patterns from a given training set so that the size of the set is reduced while its representational power is not affected. Training pattern selection is especially desirable for the effective operation of nearest-neighbour-based decision systems and for fast training of artificial neural networks. Counter-clustering is proposed here for training pattern selection. Based on a harmony-theory artificial neural network, counter-clustering exposes the clusters that exist within each class of the training set and provides a measure of the interior/exterior nature of each training pattern. Training pattern selection is, subsequently, accomplished one-shot, by retaining the boundary, isolated and exterior training patterns, while discarding most of the interior training patterns from each cluster. Examples of training pattern selection via counter-clustering are presented; the corresponding pattern classification results are reported and evaluated.