Kernel Matching Reduction Algorithms for Classification

Inspired by kernel matching pursuit (KMP) and support vector machines (SVMs), we propose a novel classification algorithm: kernel matching reduction algorithm (KMRA). This method selects all training examples to construct a kernel-based functions dictionary. Then redundant functions are removed iteratively from the dictionary, according to their weights magnitudes, which are determined by linear support vector machines (SVMs). During the reduction process, the parameters of the functions in the dictionary can be adjusted dynamically. Similarities and differences between KMRA and several other machine learning algorithms are also addressed. Experimental results show KMRA can have sparser solutions than SVMs, and can still obtain comparable classification accuracies to SVMs.