Also for k-means: more data does not imply better performance

Abstract Arguably, a desirable feature of learner is that its performance gets better with an increasing amount training data, at least in expectation. This issue has received renewed attention recent years and some curious surprising findings have been reported on. In essence, these results show more data does actually not necessarily lead to improved performance—worse even, can deteriorate. Clustering, however, subjected such kind study up now. paper shows k -means clustering, ubiquitous technique machine learning mining, suffers from the same lack so-called monotonicity display deterioration expected set sizes. Our main, theoretical contributions prove 1-means clustering monotonic, while 2-means even weakly i.e., occurrence nonmonotonic behavior persists indefinitely, beyond any sample size. For larger , question remains open.