Information Technology Journal1812-56381812-5646Asian Network for Scientific Information10.3923/itj.2010.535.544HeLiang SongQinbao ShenJunyi HaiZhen 3201093The accuracy of various instances can hardly be ensured inherently by existing k-NN prediction schemes, suffering from the single k mechanism. To address this problem, an ensemble k-NN numeric prediction algorithm, Bsk-NN, is proposed. On the basis of boosting principle, a series of base k-NN predictors is constructed firstly by Bsk-NN. Then the instance-relevant combination of each individual base predictor presents the final composite estimate. The weight of each predictor changes adaptively with respect to the distinct features of different unknown instances. Attribute selection is introduced into Bsk-NN as well to optimize the proximity measurement and to perturb the stable training set. Since the requirement that various instances demand specific prediction schemes to match with has been taken into account thoroughly, the defect of fixed k nearest-neighbor prediction is rectified consequently. Moreover, Bsk-NN is compatible with datasets of any kinds of attributes, discrete, continuous or mixed. The experimental results on public datasets show that Bsk-NN outperforms the traditional k-NN prediction and the improvement is statistically significant according to the paired t-test.]]>Breiman, L.,1999Drucker, H.,1997Freund, Y. and R.E. Schapire,1997Guyon, I. and A. Elisseeff,2003Han, E.H., G. Karypis and Y. Kumar,2001Huang, P.C.,2006Jagadish, H.V., B.C. Ooi, K.L. Tan, C. Yu and R. Zhang,2005+-tree based indexing method for nearest neighbor search.]]>Kegl, B.,2003Ridgeway, G., D. Madigan and T. Richardson,1999Schapire, R.E.,1997Schapire, R.E. and Y. Singer,1999Tahir, M.A., P. Bouridane and F. Kurugollu,2007Yamada,T., K. Yamashita, N. Ishii and K. Iwata,2006Yang, L., C. Zuo and Y.G. Wang,2005WEKA, 2008