Classification of quantized small sample data

Quantization of measured values creates a basis for data compression. In many cases the dynamics of measured values does not necessarily represent meaningful information about the underlying problem. Our aim is to study the effect of quantization on classification accuracy in small sample settings, a situation typical in microarray data classification studies. We use the equidistant quantization method and apply commonly used classifiers, namely linear discriminant analysis, linear support vector machine and k-nearest neighbor classifier. Our simulations show that data can be quantized significantly without severely hurting the classification accuracy, but using binary or ternary level data may result in significantly lower classification accuracy.