O. Dehzangi and M.Z. Jahromi (Iran)
Pattern classification, nearest neighbor, local feature weighting, adaptive distance measure
Nearest Neighbor (NN) classifier uses a distance function to predict the class of a query pattern. Several studies have shown that irrelevant, interacting, or noisy features have as much effect on distance computation as other features. Various global and local feature-weighting algorithms have been proposed to deal with this problem. In this paper, we use a local weighting scheme that allows feature weights to vary for each training instance. We propose a novel learning algorithm to learn the weight of each feature for every training instance. By associating a cost to each training instance, the proposed learning algorithm attempts to minimize the sum of costs for misclassified training instances. Using a number of data sets from the UCI-ML repository, we show that the proposed feature-weighting scheme is quite effective in improving the generalization ability of the NN classifier.
Important Links:
Go Back
