Ensemble Construction via Designed Output Distortion

摘要: A new technique for generating regression ensembles is introduced in the present paper. The based on earlier work promoting model diversity through injection of noise into outputs; it differs from methods its rigorous requirement that mean displacements applied to any data points output value be exactly zero. It illustrated how even introduction extremely large may lead prediction accuracy superior achieved by bagging. It demonstrated models with very high bias have much better than single same bias-defying conventional belief ensembling not purposeful. Finally outlined classification.