Feature selection and kernel design via linear programming
Glenn Fung, Rómer Rosales, R. Bharat Rao · 2007
The definition of object (e.g., data point) similarity is critical to the performance of many machine learning algorithms, both in terms of accuracy and computational efficiency. However, it is often the case that a similarity function is unknown or chosen by hand. This paper introduces a formulation that given relative similarity comparisons among triples of points of the form object i is more like object j than object k, it constructs a kernel function that preserves the given relationships. Our approach is based on learning a kernel that is a combination of functions taken from a set of base functions (these could be kernels as well). The formulation is based on defining an optimization problem that can be