Parallel Perceptrons, Activation Margins and Imbalanced Training Set Pruning

摘要: A natural way to deal with training samples in imbalanced class problems is prune them removing redundant patterns, easy classify and probably over represented, label noisy patterns that belonging one are labelled as members of another. This allows classifier construction focus on borderline likely be the most informative ones. To appropriately define above subsets, this work we will use base classifiers so–called parallel perceptrons, a novel approach committee machine allows, among other things, naturally margins for hidden unit activations. We shall these pattern types iteratively perform subsample selections an initial set enhance classification accuracy allow balanced performance even when sizes greatly different.