A Distributed K Nearest Neighbor Classifier for Big Data

Abstract

The
 K-Nearest Neighbor classifier is a well-known and widely applied method in data
 mining applications. Nevertheless, its high computation and memory usage cost
 makes the classical K-NN not feasible for today’s Big Data analysis
 applications. To overcome the cost drawbacks of the known data mining methods,
 several distributed environment alternatives have emerged. Among these
 alternatives, Hadoop MapReduce distributed ecosystem attracted significant
 attention. Recently, several K-NN based classification algorithms have been
 proposed which are distributed methods tested in Hadoop environment and
 suitable for emerging data analysis needs. In this work, a new distributed
 Z-KNN algorithm is proposed, which improves the classification accuracy
 performance of the well-known K-Nearest Neighbor (K-NN) algorithm by benefiting
 from the representativeness relationship of the instances belonging to
 different data classes. The proposed algorithm relies on the data class
 representations derived from the Z data instances from each class, which are
 the closest to the test instance. The Z-KNN algorithm was tested in a physical
 Hadoop Cluster using several real-datasets belonging to different application
 areas. The performance results acquired after extensive experiments are
 presented in this paper and they prove that the proposed Z-KNN algorithm is a
 competitive alternative to other studies recently proposed in the literature