We could pick up the initial k-centroids to be any of the k features in the data to be classified. But ideally, we would like to pick up the points that belong to the different clusters already in the beginning. Therefore we may want to aim to maximize their mutual distance in a certain way. Simplifying the process we could pick the first centroid to be any point from the features. The second could be the one which is furthest from the first. The third could be the one that is furthest from both first and second, and so on.