A new approach to a maximum a posteriori-based kernel classification method

Abstract

This paper presents a new approach to a maximum a posteriori (MAP)-based classification, specifically, MAP-based kernel classification trained by linear programming (MAPLP). Unlike traditional MAP-based classifiers, MAPLP does not directly estimate a posterior probability for classification. Instead, it introduces a kernelized function to an objective function that behaves similarly to a MAP-based classifier. To evaluate the performance of MAPLP, a binary classification experiment was performed with 13 datasets. The results of this experiment are compared with those coming from conventional MAP-based kernel classifiers and also from other state-of-the-art classification methods. It shows that MAPLP performs promisingly against the other classification methods. It is argued that the proposed approach makes a significant contribution to MAP-based classification research; the approach widens the freedom to choose an objective function, it is not constrained to the strict sense Bayesian, and can be solved by linear programming. A substantial advantage of our proposed approach is that the objective function is undemanding, having only a single parameter. This simplicity, thus, allows for further research development in the future.