Classification based on neural similarity

Abstract

Listen

Translate article

Following the approach of extracting similarity metrics directly from labelled data, a standard back-propagation neural network is adopted to determine a degree of similarity between pairs of input points. The similarity computed by the network is then used to guide a k-NN classifier, which associates a label with an unknown pattern based on the k most similar points. Experimental results on both synthetic and real-world data sets show that the similarity-based k-NN rule outperforms the Euclidean distance-based k-NN rule.