Classification Accuracy and Model Selection in k-Nearest Neighbors Classifiers for Data Driven Learning

Pattern classification is a core research area and a main task in pattern recognition. A classifier induced by machine learning algorithms maps an unlabeled instance to a label using internal data structures. In this paper we experiment first by changing the k value of nearest neighbors from 3 to 15 and compare the accuracy of two classifiers on various training and test sets. The results show that the attribute-weighted kNN classifier is better than others when the dataset has low dimensionality. The experimental results show that cross validation is good for improving classification accuracy with or without 3-fold cross validation for adjusting the weights. This is due to the cross validation accuracy estimator depending on the two factors of the training set and the partition into fold. Then we select nearest neighbors and change fold number from 3 to 10, the size of every fold decreases when the number of folds increase, the accuracy of classifier change, but instance-weighted kNN is relatively stable. Thus, we can see the sensitivity of k-cross validation in accuracy estimation.