The prediction accuracy of various instances can hardly be ensured by the existing k-nearest neighbor(k-NN)predictor

which works under a single k value. A combined k-NN algorithm

Bk-NN predictor

is proposed in this paper. The novel Bk-NN algorithm is used to build up a set of prediction models based on the Boosting principle

and then each model is used to predict a new instance respectively. The final predicted value of this instance is the weighted sum of these prediction values. The prediction error is reduced since the circumstance that various instances demand specific prediction models to match with is thoroughly taken into account by the Bk-NN algorithm. Moreover

the Bk-NN predictor can work well with both discrete and continuous attribute values. The experimental results on standard datasets show that

compared with the traditional k-NN predictor

the prediction accuracies of the Bk-NN predictor are improved by 6.44%-15.25%.