Fast Nearest Neighbor Classification using Class-Based Clustering

Nearest neighbor rule (NNR) is a parameter-free classifier which is easy to implement, simple to operate and with high accuracy. However, it is time and memory consuming for large datasets. This study proposed a parameter-free method to accelerate NNR. This method employs a class-based clustering algorithm to divide the training data to several clusters with respective members belonging to the same class. Cluster representations are extracted clustering border data based on the nearest neighbors between the different class clusters. Since the cluster representations are the clustering border data rather than the clustering centers, the predicting accuracy will not be affected by removing a cluster's internal data. In the predicting phase, the nearest neighbor search area is narrowed down by referring to a distance between a testing data and its nearest cluster. Thus the predicting process is speeded up. In this paper, the performance of the proposed method was evaluated and compared with NNR, K-NNR, and LIBSVM by using 5 benchmark datasets. Experimental results show that the proposed parameter-free classification algorithm is very easy to operate and gives consideration to speed and accuracy.