Anyone knows that how to do K-means clustering with significant difference? What I means is as the image shows, the area is divided into two zones using k-means clustering, and the significant difference is also obtained. I found tools that can do k-means clustering, but I have no idea how the p-value is calculated.

enter image description here

  • I'm not really sure what do you mean, but k-means converges to the local optimum (not global) so there can be a significant difference between two approaches. – dmh126 Feb 27 '19 at 9:27
  • I mean how do you know that if there is statistical significance between clusters? When the raster is classified into 2 clusters for example, maybe there is a significant difference between the two cluster, maybe there isn't. – Summer Feb 27 '19 at 9:31
  • In this case, due to its nonparametric nature, a p-value would be somewhat meaningless in regard to fit. You could however conduct a randomization test to produce model support statistics but, this is quite cumbersome with raster data. Cluster optimization, in relation to separability, is often evaluated through maximization of silhouette values. Multivariate margin distances can be scaled to probabilities using C-means but, this is at the observation level. This allows one to minimize p in finding cluster support. – Jeffrey Evans Mar 29 '19 at 15:13

Not sure if it's what you're after, but there's an R package called Jackstraw that calculates p-values for different unsupervised classifiers, including k-means. There's a paper describing it by one of the authors:

Statistical Significance of Cluster Membership

Neo Christopher Chung

January 2018

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.