Analysis of the Iterated Probabilistic Weighted K Nearest Neighbor Method, a new Distance-Based Algorithm

The k-Nearest Neighbor (k-NN) classification method assigns to an unclassified point the class of the nearest of a set of previously classified points. A problem that arises when aplying this technique is that each labeled sample is given equal importance in deciding the class membership of the pattern to be classified, regardless of the typicalness of each neighbor. We report on the application of a new hybrid version named Iterated Probabilistic Weighted k Nearest Neighbor algorithm (IPW-k-NN) which classifies new cases based on the probability distribution each case has to belong to each class. These probabilities are computed for each case in the training database according to the k Nearest Neighbors it has in this database; this is a new way to measure the typicalness of a given case with regard to every class. Experiments have been carried out using UCI Machine Learning Repository well-known databases and performing 10-fold cross-validation to validate the results obtained in each of them. Three different distances (Euclidean, Camberra and Chebychev) are used in the comparison done.