Feature subset selection for logistic regression via mixed integer optimization

This paper concerns a method of selecting a subset of features for a logistic regression model. Information criteria, such as the Akaike information criterion and Bayesian information criterion, are employed as a goodness-offit measure. The feature subset selection problem is formulated as a mixed integer linear optimization problem, which can be solved with standard mathematical optimization software, by using a piecewise linear approximation. Computational experiments show that, in terms of solution quality, the proposed method has superiority over common stepwise methods.