Evolving Vector Quantization for Classification of On-Line Data Streams

In this paper, we describe a new clustering-based classification technique (eVQ-Class), which is able to adapt old clusters and to evolve new ones on-line with new incoming data samples. It extends the conventional learning vector quantization approach, which is a kind of supervised version of original vector quantization, in mainly three points: 1.) it is able to evolve new clusters on demand by comparing new incoming samples with already generated clusters, 2.) it includes the label information in the training process by introducing a hit matrix and extending the feature space and 3.) it comes with a new weighted classification strategy. %The meaning of the concepts %'conflict' and 'ignorance' from the perception of the novel method is outlined. The novel approach will be evaluated based on high-dimensional feature data sets extracted from images recorded on-line in order to perform on-line quality control in a production process by classifying images into 'good' and 'bad' ones. The evaluation includes a comparison with well-known batch (trained and re-trained) classification techniques.