Union of Convex Separators (UCS)

We present a novel algorithm for learning a union of convex separators (UCS) to separate a class from another using decision boundaries, each of which is a convex separator. A convex separator is defined by a collection of hyperplanes that separate one class from another class using an intersection of half-spaces. A union of convex separators can be thought of as an ensemble of such convex separators that collectively separate all points of one class from the other. In this work, we put forth the notion of separability using a UCS previously known in earlier works as min-max separability, provide a gradient-based algorithm for learning UCS, and assess it against popular classifiers using recent datasets of interest.