Weightedk Nearest Neighbor Classification on Feature Projections1

H. Altay Güvenir and Aynur Akkuş Department of Computer Engineering and Information Science Bilkent University, 06533, Ankara, Turkey fguvenir, [email protected] Abstract. This paper proposes an extension to the k Nearest Neighbor algorithm on Feature Projections, called kNNFP. The kNNFP algorithm has been shown to achieve comparable accuracy with the well-known kNN algorithm. However, kNNFP algorithm has a very low time complexity compared to kNN. The extension to kNNFP introduced here assigns weights to features, therefore it is called WkNNFP, for Weighted kNearest Neighbor on Feature Projections. The paper also introduces a weight learning algorithm, called SFA, for Single Feature Accuracy. It is based on the assumption that the weight of a feature is proportional with the accuracy that will be obtained by considering only that feature. The SFA algorithm is not specific to WkNNFP, so it can be used with many other classification algorithms. An empirical evaluation of the SFA algorithm on real-world datasets shows that it achieves an important improvement in the classification accuracy of the WkNNFP algorithm.