Similarity Based Classification

摘要

We describe general conditions for data classification which can serve as a unifying framework in the study of kernel based Machine Learning algorithms. From these conditions we derive a new algorithm called SBC (for Similarity Based Classification), which has attractive theoretical properties regarding underfitting, overfitting, power of generalization, computational complexity and robustness. Compared to classical algorithms, such as Parzen windows and non-linear Perceptrons, SBC can be seen as an optimized version of them. Finally it is a conceptually simpler and a more efficient alternative to Support Vector Machines for an arbitrary number of classes. Its practical significance is illustrated through a number of benchmark classification problems.