A novel approach for discretizing continuous attributes based on tree ensemble and moment matching optimization

This paper introduces ForestDisc, an optimized, supervised, multivariate, and nonparametric discretization algorithm based on tree ensemble learning and moment matching optimization. At its core, ForestDisc uses, for each continuous attribute in the data space, moment matching to elect popular split points based on those generated while constructing a random forest model. An extensive empirical study involving 50 benchmark datasets and six classification algorithms reveals that ForestDisc is highly competitive compared with 20 major discretizers based on both intrinsic and extrinsic performance measures. The intrinsic metrics include the number of resulting bins per variable and the execution time necessary for discretizing an attribute. The extrinsic metrics concern the performance of the discretizers when applied as a preprocessing step to classification tasks, and include accuracy, F1, and Kappa measures. ForestDisc discretizer also enables an excellent trade-off between intrinsic and extrinsic performance measures.