Feature Subset Selection with TAR2less
R. Gunnalan, Tim Menzies, Kalaivani Appukutty, Amarnath Srinivasan, Ying Hu · 2003
A repeated empirical result is that machine learners can learn adequate models using a small subset of the available features. Learning from such subsets can be faster, and produces simpler models. In this paper we present a new method for feature subset selection using the TAR2 treatment learner. TAR2 assumes small backbones; i.e. a small number of features will suffice for selecting preferred classes. TAR2 can be used as a pre-processor to other learners for identifying useful feature subsets. When compared to other methods described in a recent survey by Hall and Holmes (in press), TAR2 found the smallest subsets, with minimal or no change in classification accuracy. 1.