k-means
Rafael Irizarry
To use the k-means clustering algorithm we have to pre-define , the number of clusters we want to define. The k-means algorithm is iterative. The first step is to define centers. Then each observation is assigned to the cluster with the closest center to that observation. In a second step the centers are redefined using the observation in each cluster: the column means are used to define a centroid. We repeat these two steps until the centers converge.
The kmeans function included in R-base does not handle NAs. For illustrative purposes we will fill out the NAs with 0s. In general, the choice of how to fill in missing data, or if one should do it at all, should be made with care.
Note that because the first center is chosen at random, the final clusters are random. We impose some stability by repeating the entire function several times and averaging the results.
