[http://bit.ly/k-NN] k-NN methods are closely related to Parzen windows and to kernel-based learning methods. Parzen windows use neighbourhoods of constant size (which can contain more or less than k training examples). k-NN expands or shrinks the neighbourhood to always contain exactly k training examples. Kernel-based methods are different in that they use a non-uniform neighbourhood (far-away training examples have less influence than very near examples).
