Feature Selection via Nonlinear Set Covering in Logical Analysis of Data

Abstract

In logical analysis of data, the feature selection step only considers selecting minimal number of features after binarization. This paper develops a nonlinear set covering model that explores the interaction between the number of selected original attributes and binarized features. Utilizing the partial derivative of pseudo-Boolean functions, we give a greedy heuristic for the nonlinear set covering problem. The efficacy of the algorithm is demonstrated through experiments on 10 public machine learning datasets from UCI repository.