I got this explanation of the Ward's method of hierarchical clustering from Malhotra et. al (2017), and I don't really get what it means:

Ward’s procedure is a variance method which attempts to generate clusters to minimise the within-cluster variance. For each cluster, the means for all the variables are computed. Next, for each object, the squared Euclidean distance to the cluster means is calculated. These distances are summed for all the objects. At each stage, the two clusters with the smallest increase in the overall sum of squares within cluster distances are combined.

I understand that if we have $n$ objects, then we start off with $n$ clusters. From there, they say that the means for all the variables is computed (so we got $p$ x $n$ means). And then they say that the distance from each object to to the 'cluster means' is calculated. This is where I get confused - the object means are the cluster means (at stage 1), so what distance is there to calculate?

Thanks a mil!


1 Answer 1


You are right, at the first stage, where each point $x\in X$ in your dataset $X$ is a cluster, all the distances to the cluster means are zero, so the overall sum of squared within-cluster distances is zero. So now the task is to find the two points which to combine first, such that afterward the overall sum of squared within-cluster distances is the least increased, i.e. you have to find, at stage two, the two points: $$ (x_k, x_l) = argmin_{(x_r, x_s)\in X^2} \{\|x_r - x_s\|\}. $$ You iterate as described, but in those later stages, the cluster centers are not necessarily anymore the points themselves.

  • $\begingroup$ Ah that's exactly what I thought. Just to confirm, once we've clustered some points, we then calculate the distance between other objects and its CENTROID right? $\endgroup$ Commented Apr 23, 2022 at 6:16
  • $\begingroup$ Also, isn't it funny how they then add "These distances are summed for all the objects"? I wonder what that means $\endgroup$ Commented Apr 23, 2022 at 6:17
  • 1
    $\begingroup$ @Academic005 At each stage, the $N$ points are grouped in clusters. Each of those clusters has a mean. You compute for each point the squared distance to the mean of the cluster that point is contained in. Those are $N$ squared distances that you sum up. That's what they refer to when saying: "These distances are summed for all the objects". $\endgroup$
    – frank
    Commented Apr 23, 2022 at 14:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.