Let's say I've created some clusters with Kmeans using 5 features, the Silhouette Score for these cluster are very high, higher than 0.8, and The within-cluster sum of squares is around 130 in this example.

Now only looking at cluster A, the average of a feature X inside the cluster is 100. My issue is that there are a lot of examples in which this metric has the value of 10, or 1000, or 4000, basically the examples inside the cluster are very different. I'm assuming they are falling in the same cluster because between the 5 features (one of them being the feature X I referred to) they have the same value for one of this features, but are very different when it comes to the other 4 features.

I see this as a problem because I can't say these examples are similar just because they have 1 or 2 features in common and are very different in the other features.

I've tested multiple K's but there's still a lot of variance inside these clusters. I'm also measuring the WSS (The within-cluster sum of squares) to try and minimize it as much as possible assuming it would help me avoid that many different examples.

I've not seen any content regarding this type of issue on-line.

Is this normal in Kmeans? Am I wrong in thinking most things inside my cluster should look alike? How can I avoid this?

  • $\begingroup$ This is a more general problem with assuming a method off-the-shelf will fit a problem. One option is to define a custom distance between points that encourages feature X to be more important—warping the way you measure the feature space. $\endgroup$ May 14 at 20:39


Your Answer

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

Browse other questions tagged or ask your own question.