Given a collection of observatons $\{X_i\}_1^N$ and prespecify the number of clusters K. The K-means solves $$ \underset{\{C_k\}_1^K}{\arg\min} \sum_{k=1}^K \sum_{i \in C_k}|| X_i - \mu_k||^2 $$ where $||\cdot||$ is the $\ell_2$ norm, $\mu_k$ is the mean of $X_i$ in $C_k$. In ISL and R function kmeans(), $\sum_{i \in C_k}|| X_i - \mu_k||^2$ is the within-cluster sum of square.

My questions:

Q1. AS the value of K increases, what is the behavior of the total within-cluster sum of squares $\sum_{k=1}^K \sum_{i \in C_k}|| X_i - \mu_k||^2$?

Q2. (This is an open question) In order to increase the accuracy of assignment, I attempt to choose a sufficiently large K. I hope that each subgroups will only(mostly) contain observations from the common true group. Then we can combine these subgroups to approximate the unknow true clustering. Is my idea feasible?

My thinking:

Q1. I see that kmeans is a random algorithm and it is likely to find a locally minima (discussed here). But is there any theorem(paper, or case study) to explain the (stochastic) behavior of total within-cluster sum of squares and its upper bound (or other results)?

Take the iris data in R as an example enter image description here

Q2. I plot the partition result for K varying form 3 to 12. Different colors(black, blue and red) represent the true species of each observation, and different symbols represent different clusters from kmeans. It seems that my idea is meaningful.




Thank you for your opinions!

  • 1
    $\begingroup$ For each value of k, there exists the corresponding global optimum - the utterly minimal SSwithin. This minimum is, of course, diminishes as k grows. No proving is necessary (albeit it can be provided and you might find it on this site), only common sense. But the global optimum is not necessarily found with a particular data and particular k. $\endgroup$
    – ttnphns
    Sep 1, 2021 at 9:55
  • 1
    $\begingroup$ See e.g. stats.stackexchange.com/q/51805/3277, stats.stackexchange.com/q/159166/3277 and other threads $\endgroup$
    – ttnphns
    Sep 1, 2021 at 9:58
  • $\begingroup$ Q2. The principal way to get the solution equal or close to the global optimum is not changing k (though it is trivial that as k approaches n, the global optimum is easier to find, in less iterations) but to get good estimates of initial centroids. There are methods. $\endgroup$
    – ttnphns
    Sep 1, 2021 at 10:06
  • $\begingroup$ @ttnphns Thank you! I once hoped to specify a large enough initial K so that the algorithm can approach the global optimum at any initial centroids. Thank you for your helpful opinion! I will reconsider the initial centroids selection. $\endgroup$
    – HJ Liang
    Sep 1, 2021 at 10:42

1 Answer 1


Preamble: You talk about the "underlying true clusters", but in applied clustering this is a highly problematic concept. Assuming a certain model, one can define what is meant by "true clustering", but more than one definition is possible (for example a mixture distribution of 6 Gaussians may have only three modes, and one can define the "true clustering" as corresponding to the six Gaussians or the three modes, which gives different concepts for what the term "true clustering" actually means). In a real application, it will always depend on the nature of the data and the aim of clustering what kind of "true clusters" a researcher wants to find. The data on their own can not decide this for you! Some more insight is in this excerpt from the Handbook of Cluster Analysis: https://arxiv.org/abs/1503.02059

Another thing to know here is that the behaviour of the global optimum of the K-means objective function may differ from the behaviour of local optima that an algorithm actually will find. Unfortunately, unless the data set is very small, no algorithm that runs in realistic time will guarantee that you find the global optimum.

Q1: Using the global optimum, the WSS (within-cluster sum of squares) will never increase with increasing $K$ (this can be proved showing that if you have a solution for $K-1$ clusters, you can always reproduce the same WSS value for $K$ clusters not assigning any point to the added $K$th cluster, keeping all the other cluster centers as they are, so the same WSS can always be achieved - in fact, as long as there are at least $K$ distinct points, one can even easily always decrease the WSS by putting cluster center $K$ exactly on a point where no cluster center was before).

This does not necessarily hold for a local optimum in a real analysis. However, it is probably safe to say that the WSS should mostly decrease when increasing $K$ even then. It is also possible to construct an algorithm for finding local optima so that the WSS cannot increase when increasing $K$, but this may not be the best algorithm to be had when looking at other performance criteria. Furthermore, if $K=N$ (in fact if $K$ equals the number of distinct points in the data), the WSS will be 0, its lowest possible value.

Q2: The preamble implies that any answer to this question requires the nontrivial and non-unique definition of what you actually mean by "true clusters", and in any real application, neither will the data tell you this, nor is there any guarantee that any formal model-based definition will correspond to what is really meaningful. No real progress (as in: "really meaningful for real data") can be made here without knowledge of what the data are, the aim of clustering, and potentially other background of the real application.

That said, your idea is not stupid, and there is some literature about it:



Actually, the traditional Ward's hierarchical clustering method can be interpreted as doing a very similar thing (at each stage two clusters are combined by use of the WSS criterion). However, as said before, there is no way to determine from the data alone without additional user input how much "combining" there should be and when to stop.

  • $\begingroup$ Thank you! Your detailed explanation of "true clusters" deepens my understanding. I think, in statistical modeling for heterogeneous data, firstly we assume(of course this assumption may be wrong, then we need to build other models to fit the data) the true data generating process of the data follows a group-specific model, eg, the group-specific linear model. Secondly, in model estimation, we need to group the data and then estimate the group-specific slopes. Therefore, in this sense, we can define the true clusters from the model, if the model is appropriate to the data. Am I right? $\endgroup$
    – HJ Liang
    Sep 1, 2021 at 11:28
  • $\begingroup$ Let's say this is a standard approach of doing it, so surely not wrong although not the only thing that can be done. However, if you talk about linear models, K-means seems to be an inappropriate starting point, as in linear models the $x$- and $y$-variables are not symmetric whereas K-means treats all variables symmetrically. You may want to google for clusterwise regression or regression mixtures if that's what you are dealing with. $\endgroup$ Sep 1, 2021 at 13:17
  • $\begingroup$ ok, I see the difference. Thank you for your helpful suggestions! $\endgroup$
    – HJ Liang
    Sep 1, 2021 at 13:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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