An evolutionary cut points search for graph clustering-based discretization

Most discretization algorithms focus on the univariate case, which proceeds without considering interactions among attributes. Furthermore, they find the value of discretization criterion based on a greedy method that usually leads to a sub-optimal set of cut-points. In response, this paper introduces an evolutionary cut points search for graph clustering-based discretization (GraphE). The resulting method exhibits both multivariate and searching properties. The proposed evolutionary model is based on Genetic algorithm with the specifically designed fitness function. It simultaneously considers the data similarity via the notion of normalized association and the number of intervals. The proposed method is compared with 7 state-of-the-art discretization algorithms, conducted on 15 datasets and 3 classifier models. The results suggest that the new technique usually achieves higher classification accuracy than the comparative methods, while requiring less computational time than the existing optimization-based model.