EKSL: An effective novel dynamic ensemble model for unbalanced datasets based on LR and SVM hyperplane-distances

Abstract

Unbalanced data is considered in many real-world classification problems, where it is often costly in practice to sample and establish a homogeneous class distribution for a minority class. The choice of methods, diversity of datasets used for structuring, and the correct kernel decision are quite decisive in the success of the system. This study develops a powerful classifier algorithm that provides an alternative solution to kernel experiments with the choice of a general-purpose, fast, automatic linear kernel. Principally based on SVM, k-means clustering in partitions is used, and logistic regression is integrated into the ensemble system. To increase the success rate and deal with the maximum convergence problem, the soft margin value of the standard SVM is changed in an adaptive structure. In the experiments, it has been observed that accuracy performance showed an improvement of (1–8) %, even when the optimal k value in the k-means clustering stage of the proposed ensemble methodology has been compared to the best in other kernels measurements. In particular, when the unbalance property of the data gets an increase, it has been observed thatbetter results are obtained in all evaluation metrics than in a single classifier.