A Skew-tolerant Strategy and Confidence Measure for k-NN Classification of Online Handwritten Characters

Confidence measures for k-NN classification are an important aspect of building practical systems for online handwritten character recognition. In many cases, the distribution of training samples across the different classes is marked by significant skew, either as a consequence of unbalanced data collection or because the application itself incrementally adds samples to the training set over a period of use. In this paper, we explore the adaptive k-NN classification strategy and confidence measure in the context of such skewed distributions of training samples, and compare it with traditional confidence measures used for k-NN classification as well as with confidence transformations learned from the data. Our experiments demonstrate that the adaptive k-NN strategy and confidence measure outperforms other measures for problems involving both large and small sets of training data.