DG-subspace: A novel attributes selection method for lazy learning

Lazy learning has shown promising reliability in data stream classification mining, which suffer from `Curse of dimensionality´ in broad applications. Conventional Attribute selection methods always seek promising subspace by ranking all the attributes, which is not suitable for lazy learning, and suffer from high computing complexity. We proposed a novel attributes selection method `DistinGuishing Subspace (DG-Subspace)´, which lay high values on the performance of attributes as a group instead of single attribute with higher ranks. `DistinGuishing Pattern Tree (DGP-tree)´ was formed to compress dataset, based on which a heuristic method to seek DG-subspace was raised, with linear scalability. Theoretic analysis and numeric experiment justified the effectiveness and efficiency of the method.