Dynamic k determination in k-NN classifier: A literature review

One of the widely used classification algorithms is k-Nearest Neighbours (k-NN). Its popularity is mainly due to its simplicity, effectiveness, ease of implementation and ability to add new data in the training set at any time. However, one of its main drawbacks is the fact that its performance is highly dependent on the proper selection of parameter k, i.e. the number of nearest neighbours that the algorithm examines. The most frequently used technique for the “best” k determination is the cross validation as there is no general rule for choosing the k value due to its dependency on the training dataset. However, selecting a fixed k value throughout the dataset does not take into account its special features, like data distribution, class separation, imbalanced classes, sparse and dense neighborhoods and noisy subspaces. A lot of research has been done to date in the specific field, leading to many k-NN variations. In the present research, a thorough literature review is conducted in order to summarize all the achievements made to date in this field. Specifically, a pool of twenty eight (28) approaches with their experimental results are presented, all concerning methods and techniques for dynamic “best” k selection.