The performance of Nearest Neighbor (NN) classifier is highly dependent on the distance (or similarity) function used to find the NN of an input test pattern. Many of the proposed algorithms try to optimize the accuracy of the NN rule using a weighted distance function. In this scheme, a weight parameter is learned for each of the training instances. The weights of training instances are used in the generalization phase to find the NN of an input test pattern. The Weighted Distance Nearest Neighbor (WDNN) algorithm attempts to maximize the leave-one-out classification rate of the training set by adjusting the weight parameters. The procedure simply leads to weights that overfit the train data, which degrades the performance of the method especially in noisy environments.
In this paper, we propose an enhanced version of WDNN, called Overfit Avoidance for WDNN (OAWDNN), that significantly outperforms the algorithm in generalization phase. The proposed method uses an early stopping approach to decrease instance weights specified by WDNN, which implicitly makes the class boundary smooth and consequently more generalized.
In order to evaluate robustness of the algorithm, class label noise is added to a variety of UCI datasets. The experimental results show the supremacy of the proposed method in generalization accuracy.
Keywords
