The Use of k-Means Algorithm to Improve Kernel Method via Instance Selection

摘要

The kernel method is well known for its success in solving the curse of dimension of linearly inseparable problems. But as an instance-based learning algorithm it suffers from high memory requirement and low efficiency in that it needs to store all of the training instances. And when there are noisy instances classification accuracy can suffer. In this paper we present an approach to alleviate both of the problems mentioned above by using k-means algorithm to select only k representativeness instances of the training data. And we view the selected k instances as the new data set, where the choice of the value of k is influenced by the size and the character of the data set. It turn out that with a carefully selected k we can still get a good performance while the number of the instances stored are greatly decreased.