Compact MILP models for optimal and Pareto-optimal LAD patterns

This paper develops MILP models for various optimal and Pareto-optimal LAD patterns that involve at most 2n2n 0–1 decision variables, where nn is the number of support features for the data under analysis, which usually is small. Noting that the previous MILP pattern generation models are defined in 2n+m2n+m 0–1 variables, where mm is the number of observations in the dataset with m≫nm≫n in general, the new models are expected to generate useful LAD patterns more efficiently. With experiments on six well-studied machine learning datasets, we first demonstrate the efficiency of the new MILP models and next use them to show different utilities of strong prime patterns and strong spanned patterns in enhancing the overall classification accuracy of a LAD decision theory.