Why are there different clusters when we initialize other points to be the cluster separator?
I thought there must be an ideal clustering that the system would converge to. How can these different clusterings be justified?
Hi @yusufnzm,
welcome to the community and thanks for your question!
Why are there different clusters when we initialize other points to be the cluster separator?
For different initialisations you might end up in different local optima. Besides initialisation also randomness of the training processes can contribute to this, see also this thread: How different initialization of centroids of K-means results in drastic different clusters ? They all share common cost function - #4 by Christian_Simonis
A criterion how clustering can be justified or assessed is e.g. (Shannon) entropy-based criteria like Adjusted Mutual Information, see also further clustering criteria here: Performance Metrics in ML - Part 3: Clustering | Towards Data Science
If you are uncertain what a good number of clusters is, you can take a closer look at the elbow method which is often used in practice.
Hope that helps!
Best regards
Christian
1 Like
Thank you, man. Very appreciated.