"Feature selection for multi-class discrimination via mixed-integer linear programming", F. J. Iannarilli, Jr. and P. A. Rubin, IEEE Trans. Pattern Analysis and Machine Intelligence 25:6 (2003).

We reformulate branch-and-bound feature selection employing L_∞ or particular L_p metrics as mixed-integer linear programming (MILP) problems, affording convenience of widely available MILP solvers. These formulations offer direct influence over individual pairwise interclass margins, which is useful for feature selection in multi-class settings.