Simple Power Series for Pattern Classification

Abstract

We show in this paper, that a simple power series model with appropriate learning formulation, can be used for effective pattern classification. Essentially, an error counting cost function is adopted. Through a linear parametric power series model and a quadratic approximation to the error cost, a deterministic solution is derived. This solution is seen to relate to a class-specific setting of the more generic weighted least-squares. A tuning mechanism is thus incorporated for robust applications. Our empirical evaluations show effectiveness of the classifier.