Feature Selection using e-values

Navigating an exponentially growing feature space using wrapper methods is In the context of supervised parametric models, NP-hard (Natarajan, 1995), and case-specific search strategies we introduce the concept of e-values. An e-value like k-greedy, branch-and-bound, simulated annealing is a scalar quantity that represents the proximity of are needed. Sparse penalized embedded methods can tackle the sampling distribution of parameter estimates high-dimensional data, but have inferential and algorithmic in a model trained on a subset of features to that issues, such as biased Lasso estimates (Zhang & Zhang, of the model trained on all features (i.e. the full 2014) and the use of convex relaxations to compute approximate model). Under general conditions, a rank ordering local solutions (Wang et al., 2013; Zou & Li, 2008). of e-values separates models that contain all essential features from those that do not. Feature selection in dependent data models has received comparatively lesser attention. Existing implementations of The e-values are applicable to a wide range of wrapper and embedded methods have been adapted for dependent parametric models. We use data depths and a fast data scenarios, such as mixed effect models (Meza resampling-based algorithm to implement a feature & Lahiri, 2005; Nguyen & Jiang, 2014; Peng & Lu, 2012) selection procedure using e-values, providing and spatial models (Huang et al., 2010; Lee & Ghosh, 2009).