I understand that Certroid-based classifiers are better than k-NNs.
220.127.116.11 Centroid-Based Classiﬁcation
Centroid-based classiﬁcation is a fast alternative to k-nearest neighbor classiﬁers. The basic idea is to use an oﬀ-the-shelf clustering algorithm to partition the documents of each class into clusters. The number of clusters derived from the documents of each class is proportional to the number of documents in that class... ... ... ... ..., this approach also indirectly addresses the issues of synonymy and polysemy, with the additional advantage that the k-nearest neighbor classiﬁcation can be performed more eﬃciently with a smaller number of centroids. The dominant label from the top-k matching centroids, based on cosine similarity, is reported. Such an approach can provide comparable or better accuracy than the vanilla k-nearest neighbor classiﬁer in many cases.
This is quoted from "Data Mining: The Textbook" by Charu C. Aggarwal. Page-447-48.
But, When shouldn't we use Centroid-based classifiers as opposed to Nearest Neighbors?
My teacher told me something about Biases which I don't remember well.