Predicting the benefit of sample size extension in multiclass k-NN classification

In industrial quality inspection obtaining the training data needed for classification problems is still a very costly task. Nevertheless, the classifier quality is crucial for economic success. Thus, the question whether the influence of the training data on the classification error has been fully exploited and enough data has been obtained is very important. This paper introduces a method to answer this question for a specific problem. To be able to make a concrete statement and not only general recommendations, we focus on the k-NN classifier, since it is widely used in industrial implementations. The method is tested on four different multiclass problems: original data from an optical media inspection problem, the MNIST database, and two artificial problems with known probability densities.