Automated feature selection with a distinction sensitive learning vector quantizer

Abstract

An extended version of Kohonen's Learning Vector Quantization (LVQ) algorithm, called Distinction Sensitive Learning Vector Quantization (DSLVQ), is introduced which overcomes a major problem of LVQ, the dependency on proper pre-processing methods for scaling and feature selection. The algorithm employs a weighted distance function and adapts the metric with learning. Highest weights are assigned to components in the input vectors which are most informative for classification; non-informative components are discarded. The algorithm is applied to the analyses of multi-channel EEG data and compared with experienced methods.