Construction of convex hull classifiers in high dimensions

► Proposed is an algorithm to find a small number of approximated convex hulls enclosing all training samples of each class. ► Classification is made by the closest convex hull and its class label. ► An advantage of the proposed algorithm is to maximize non-linear margin in the original feature space unlike SVM. ► Another advantage is that it is naturally able to deal with multi-class problems. ► One drawback is the high cost of constructing the approximate convex hulls in high dimensions.